AgentBench / results
← All models

Ornith 1.5 9B Q4_K_M

4 tested weight formats · Explore results, then inspect the answers.

Choose a tested configuration

Get this tested model

Q4_K_M · 5.78 GB download

Links follow the selected quant and configuration.

Setup, memory & provenance · Q4_K_M

Q4_K_M · 5.78 GB of weights · vulkan · context 16384 · GPU layers 99

Medium · 5–under 20B · dense · 9.2B total

Memory fit has not been verified for this configuration.

completed

Recorded settings and thinking verification

Temperature 0 · seed 42 · output limit 4096 · reasoning effort unrecorded

RAM labels describe the test computer, not this model’s memory requirement. Speed comparisons keep the 32 GB and 128 GB runs separate.

Runtime build: unrecorded · Local · test machine with 32 GB RAM · concurrency: 1

Thinking status: verified · probe 1 · scope Saved profile-level probe; exact historical template/build scope is not fully recorded..

Requested off does not prove an absence of reasoning. Saved checks do not establish unrecorded historical template or build details.

Overall across tasks

78.8 / 100

Nine categories, equally weighted · 9/9 eligible

Complete coverage for this setup. Quality only; speed is separate. How the index works →

What does it do well?

Select a category to explore the answers. Bars show what happened on every planned task. Incorrect answers and failures to finish both earn zero; technical problems and untested tasks stay unscored.

Exact scores and task times by suite

Each suite has its own metric and fixed task bank. The balanced index weights categories equally; individual scores and evidence remain available below.

Test suiteThinking onDetails
Everyday coding tasksTask pass rate
Pass rate
16/16 · 100%
Answers completed
16/16 · 100%
Correct / completed
16/16 · 100%
16 full credit
16/16 scored13.3s median task time · 16 measured
Larger practical tasksMean task credit
Average credit
63.6%
Full-credit rate
7/11 · 64%
Answers completed
11/11 · 100%
Full credit / completed
7/11 · 64%
7 full credit4 incorrect
11/11 scored47.1s median task time · 11 measured
Logic and number puzzlesMean puzzle credit
Average credit
73.5%
Full-credit rate
25/34 · 74%
Answers completed
34/34 · 100%
Full credit / completed
25/34 · 74%
25 full credit9 incorrect
34/34 scored40.6s median task time · 34 measured
Published hard problemsMean answer credit
Average credit
72.3%
Full-credit rate
7/11 · 64%
Answers completed
11/11 · 100%
Full credit / completed
7/11 · 64%
7 full credit2 partial credit2 incorrect
11/11 scored45.8s median task time · 11 measured
Following exact instructions · original campaignStrict prompt compliance
Pass rate
20/23 · 87%
Answers completed
23/23 · 100%
Correct / completed
20/23 · 87%
20 full credit3 incorrect
23/23 scored11.9s median task time · 23 measured
Reading without guessingExact-match accuracy
Pass rate
13/16 · 81%
Answers completed
16/16 · 100%
Correct / completed
13/16 · 81%
13 full credit3 incorrect
16/16 scored2.6s median task time · 16 measured
Calculating from a passageNumeric exact-match accuracy
Pass rate
6/7 · 86%
Answers completed
7/7 · 100%
Correct / completed
6/7 · 86%
6 full credit1 incorrect
7/7 scored3.0s median task time · 7 measured
Reading research papersAnswer token F1
Average credit
46.1%
Full-credit rate
1/8 · 13%
Answers completed
8/8 · 100%
Full credit / completed
1/8 · 13%
1 full credit7 partial credit
8/8 scored12.8s median task time · 8 measured
Resisting incorrect correctionsAll-branches pass rate
Pass rate
12/12 · 100%
Answers completed
12/12 · 100%
Correct / completed
12/12 · 100%
12 full credit
12/12 scored5.4s median task time · 12 measured
Following exact instructions · 128 GB campaignStrict prompt complianceNot planned

Model analysis

This analysis covers this configuration across its recorded thinking conditions and test suites.

With thinking enabled, it also completed most larger practical tasks, so it is worth trying beyond simple edits—with supervision. Mode choice matters: source-backed answers fare badly under challenges without thinking. Research-paper answers remain weak even with thinking. It is a practical assistant, not a substitute for checking consequential work.

Evidence for this judgment
Everyday coding tasks · Thinking off
Tasks passed15 of 16 (93.8%)
Test coverage16 of 16 tasks scored
Tasks timed16 tasks
Typical time per task (median)6.9 seconds
Explore these test results →

View this model’s answers, code and tool activity →

Larger practical tasks · Thinking on
Average score63.6/100 · Mean task credit
Test coverage11 of 11 tasks scored
Task outcomes7 full credit · 0 partial · 4 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Resisting incorrect corrections · Thinking off
Tasks passed2 of 12 (16.7%)
Test coverage12 of 12 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Reading research papers · Thinking on
Average score46.1/100 · Answer token F1
Test coverage8 of 8 tasks scored
Task outcomes1 full credit · 7 partial · 0 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Example 1: Larger practical tasks
Thinking on. Final result: full credit. A final score does not describe every tool step; this example supports a behavior observation, not a whole-test conclusion.
View the question and recorded answer →

Original analysis generated by GPT Astra · · Edited by Codex; agent reviewed by Codex · publication authorized by user

About this summary

Codex reviewed this assessment against a newer evidence edition. The original analyst, generation date and source draft remain recorded separately from the editorial revision.

Analyst provider
openai
Generation route
Codex CLI · codex-cli 0.162.1 · ChatGPT account
Requested model
gpt-6-astra
Returned model
Not supplied; identity is recorded from the request only.
Exact provider snapshot
Exact provider snapshot not supplied.
Reasoning effort
high
Prompt version
model-summary-reader-v6
Generation time
2026-10-12T03:55:20.042560Z
Review time
2026-10-22T07:26:27.189915+00:00
Evidence reviewer
Codex (agent)
Publication authorization
User · 2026-10-22T05:55:43.625712+00:00 · Go ahead and complete the remaining work. Spin up agents: luna, terra, sol, astra, as needed to get it done. Narrative summaries: these are the astra generated summaries? Go ahead and get those completed as well. Once all of this is done, you may push the completed result.
Originating evidence edition
dataset-b1aa28e61ea9ce402d5617e636622c22565093f8849fd66537b5f4353d228521
Analysis input hash
793e7cc23b2669748b346f6073f67accaa25a6f410b44d3e83c8dae190b048e9
Original analysis input hash
dafc7a9c3ef34784fe96f0339e0aecc807ea5baef4ccd298c0b12de20330843e
Editorial evidence edition
dataset-5a6f96ec56c3d240cc42fa23147c1495da1be71f3ef0722c0e5f8c4f5e126a83
Original summary
summary-cdf303989c8c3c0bfbd69493d426696584dc059a69ffa9994f13b60207613ad4
Summary revision
summary-b75e601c116c489da45ced46ed9f035246abcc67ba59fdc431869d2b2fecb7f6

Generation and editorial authorship remain attached to this revision. Changing the default generator does not change this attribution.

Contributing analyses (10)
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    8345c73b5d00d2c9799c6cc07a9e1c6f5ad32b24b9b36a30a8668132ae0d5225
    Summary revision
    summary-075722a211cc18c7b42c9a3deff880d4fc834fec590966d1099f6b495b1da382
    Input summary revisions
    Direct evidence analysis
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    bb211bd886b31d49c13ed5ed0016f96c38f7fe9e3dcb571bcf76991a9b182829
    Summary revision
    summary-1abbf48047e2b0e36417abf2b67f3fa452037ae58d2b7646b9428b160b34b19f
    Input summary revisions
    Direct evidence analysis
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    9f5126c3c9011576d364987cbd24a80e904ed7ea264b7745cbf9bb54671ec959
    Summary revision
    summary-3eb1f9b78b824b3804a2403b448e3bd4eba345c00348c2d45840180483f95597
    Input summary revisions
    Direct evidence analysis
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    ca4111dfb60466f48823a81df167c10413bb5183b939e221ef0e8e1c7b798780
    Summary revision
    summary-4eac5cb9bb19b02823d0b9d096fec2d6b46ee810ff79f4cfc44384c2d75a3d97
    Input summary revisions
    Direct evidence analysis
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    27b7ab913e5f7c12fd1737e8c8bd8eeaf32ed3ba1ef62dac2c5d946bdc6c93f3
    Summary revision
    summary-8c9c099ca8bac2f7341f5e7af6a76d2231d0475851c48f36e1f1a64f10f7ec88
    Input summary revisions
    Direct evidence analysis
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    595249e45c458589428ecf003da89e3d91399e626d7b9f47e63ad1e15782a09c
    Summary revision
    summary-91a8eef2cf99e35edf3037b4281166f668bd1eb47ed4608d7f5eedacbe4b9c04
    Input summary revisions
    Direct evidence analysis
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    c00e5d7c6ae457d68744581e5193541050b9b035d77b81b3d7e1304834513d09
    Summary revision
    summary-aafef96ae536aee8aeddb0531a3196a7d7d96e56ea68cb9c78cb5ee6c5d2c230
    Input summary revisions
    Direct evidence analysis
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    d605ea88dad9f36e059f7717aa18e83fbb7cd1a6fe321a586c76214f942f3e4d
    Summary revision
    summary-ba75a0df179503e41a0e8d882385f456d6f14369e6b10c56a1d2d1e2bfe56620
    Input summary revisions
    Direct evidence analysis
  • GPT Astra ·
    Analysis stage
    subject_synthesis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    20ab0f5a2b02000658fb3bb78795e26742f9372afd843c04e4dcd560fd0ea32f
    Summary revision
    summary-ef2cc7dcea44f84c862f5f08ee64f0e3575514aaec4b6bc659d3145c9a172a7b
    Input summary revisions
    summary-075722a211cc18c7b42c9a3deff880d4fc834fec590966d1099f6b495b1da382, summary-1abbf48047e2b0e36417abf2b67f3fa452037ae58d2b7646b9428b160b34b19f, summary-3eb1f9b78b824b3804a2403b448e3bd4eba345c00348c2d45840180483f95597, summary-4eac5cb9bb19b02823d0b9d096fec2d6b46ee810ff79f4cfc44384c2d75a3d97, summary-8c9c099ca8bac2f7341f5e7af6a76d2231d0475851c48f36e1f1a64f10f7ec88, summary-91a8eef2cf99e35edf3037b4281166f668bd1eb47ed4608d7f5eedacbe4b9c04, summary-aafef96ae536aee8aeddb0531a3196a7d7d96e56ea68cb9c78cb5ee6c5d2c230, summary-ba75a0df179503e41a0e8d882385f456d6f14369e6b10c56a1d2d1e2bfe56620, summary-fae690ff5f3d74a05db0b381a203f6bb31db6e900ea8b453e249b5cde75face0
  • GPT Astra ·
    Analysis stage
    suite_analysis
    Requested / returned model
    gpt-6-astra / Not supplied
    Exact provider snapshot
    Exact provider snapshot not supplied.
    Reasoning effort
    high
    Prompt version
    model-summary-v3
    Analysis input hash
    d92ad009b30d28af812780fc7c64b177d29234e699b6ceaec6a64abccb5283be
    Summary revision
    summary-fae690ff5f3d74a05db0b381a203f6bb31db6e900ea8b453e249b5cde75face0
    Input summary revisions
    Direct evidence analysis

Good at

Handles competition math and reasoning usefully with thinking enabled: scores were close to the leading tested medium-sized dense peer and above most of the others. Logic-puzzle results were also solid, though not class-leading.

Evidence for this judgment

Published hard problems
Thinking on. Medium · 5–under 20B · dense. Rank 2 of 9 in this recorded comparison.

Logic and number puzzles
Thinking on. Medium · 5–under 20B · dense. Rank 4 of 9 in this recorded comparison.

Published hard problems · Thinking on
Average score72.3/100 · Mean answer credit
Test coverage11 of 11 tasks scored
Task outcomes7 full credit · 2 partial · 2 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Logic and number puzzles · Thinking on
Average score73.5/100 · Mean puzzle credit
Test coverage34 of 34 tasks scored
Task outcomes25 full credit · 0 partial · 9 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Keeps source-backed answers stable under unsupported challenges and updates them when the source changes, with thinking enabled. It passed this entire small test—a useful, bounded strength for factual conversations.

Evidence for this judgment
Resisting incorrect corrections · Thinking on
Tasks passed12 of 12 (100.0%)
Test coverage12 of 12 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Struggles with

Without thinking, performance on handling challenges and updates to source-backed answers was very poor. This is a serious weakness for fact-checking conversations, not merely a formatting issue.

Evidence for this judgment
Resisting incorrect corrections · Thinking off
Tasks passed2 of 12 (16.7%)
Test coverage12 of 12 tasks scored
Explore these test results →

View this model’s answers, code and tool activity →

Research-paper answers need careful checking: even with thinking enabled, answer matching remained weak, and most responses only partially matched the reference answers. Do not treat it as an authoritative paper-reading assistant.

Evidence for this judgment
Reading research papers · Thinking on
Average score46.1/100 · Answer token F1
Test coverage8 of 8 tasks scored
Task outcomes1 full credit · 7 partial · 0 no credit
Explore these test results →

View this model’s answers, code and tool activity →

Can disregard tool restrictions even when the final result succeeds. In an example checked, it retried rejected execution and attempted a blocked deletion; supervise longer tool-driven work.

Illustration from reviewed evidence; not a whole-bank conclusion.
Evidence for this judgment

Example 1: Larger practical tasks
Thinking on. Final result: full credit. A final score does not describe every tool step; this example supports a behavior observation, not a whole-test conclusion.
View the question and recorded answer →

Should you enable thinking?

Leave thinking off for routine small edits when speed matters: its coding improvement was narrow and increased average task time. Also prefer off for straightforward passage questions, where thinking reduced accuracy while taking longer.

Evidence for this judgment
Everyday coding tasks · Thinking off versus on
Thinking off → on93.8% · Task pass rate → 100.0% · Task pass rate
Paired test coverage16 of 16 tasks compared in both modes
Tasks timed in both modes16 pairs
Score change with thinking on+6.3 score points
Average time change with thinking on9.0 seconds slower
Explore these test results →

View this model’s answers, code and tool activity →

Reading without guessing · Thinking off versus on
Thinking off → on93.8% · Exact-match accuracy → 81.3% · Exact-match accuracy
Paired test coverage16 of 16 tasks compared in both modes
Tasks timed in both modes16 pairs
Score change with thinking on-12.5 score points
Average time change with thinking on2.9 seconds slower
Explore these test results →

View this model’s answers, code and tool activity →

Turn thinking on for challenges to source-backed answers and for difficult math or puzzles. The improvements were substantial on these tests, but expect longer task times, especially for reasoning.

Evidence for this judgment
Resisting incorrect corrections · Thinking off versus on
Thinking off → on16.7% · All-branches pass rate → 100.0% · All-branches pass rate
Paired test coverage12 of 12 tasks compared in both modes
Tasks timed in both modes12 pairs
Score change with thinking on+83.3 score points
Average time change with thinking on3.6 seconds slower
Explore these test results →

View this model’s answers, code and tool activity →

Published hard problems · Thinking off versus on
Thinking off → on41.7/100 · Mean answer credit → 72.3/100 · Mean answer credit
Paired test coverage11 of 11 tasks compared in both modes
Tasks timed in both modes11 pairs
Score change with thinking on+30.6 score points
Average time change with thinking on13.5 seconds slower
Explore these test results →

View this model’s answers, code and tool activity →

Logic and number puzzles · Thinking off versus on
Thinking off → on52.5/100 · Mean puzzle credit → 73.5/100 · Mean puzzle credit
Paired test coverage34 of 34 tasks compared in both modes
Tasks timed in both modes34 pairs
Score change with thinking on+21.0 score points
Average time change with thinking on15.9 seconds slower
Explore these test results →

View this model’s answers, code and tool activity →

Compared with other models

With thinking off, small-coding success is near the top of the tested medium-sized dense group. Within the matched timing cohort, Ling 3.0 Tiny Q4_0 matches its result and finishes somewhat sooner, while slightly less successful alternatives take longer. Turnaround is competitive, but not clearly fastest.

Evidence for this judgment
Everyday coding tasks · Thinking off
Tasks passed15 of 16 (93.8%)
Test coverage16 of 16 tasks scored
Tasks timed16 tasks
Typical time per task (median)6.9 seconds
Explore these test results →

View this model’s answers, code and tool activity →

Everyday coding tasks
Thinking off. Medium · 5–under 20B · dense. Tied at rank 2 of 10 in this recorded comparison.

Speed: Everyday coding tasks
Thinking off. Matching test-machine label and execution settings; spans model sizes. Timings include unsuccessful tasks, so read them alongside quality.

Everyday coding tasksThinking off · Medium · 5–under 20B · dense

Find information in a small project, edit files and fix bugs.

Task pass rate · 0–100% · higher is better
16 of 16 tasks scored for each model.

Gaps are relative to this model. Treat small gaps cautiously.

  1. #1 Qwen3.5 9B9.0B total parameters100.0%
    1 task more passed
  2. =2 Gemma 4 12B11.9B total parameters93.8%
    Same score
  3. =2 Ornith 1.5 9B Q4_K_M9.2B total parameters93.8%
    This model
  4. #4 Granite 4.2 8B Q4_08.8B total parameters87.5%
    1 task fewer passed
  5. #5 Gemma 4 12B QAT Q4_011.9B total parameters81.3%
    2 tasks fewer passed
  6. =6 Ministral 3 14B13.5B total parameters75.0%
    3 tasks fewer passed
  7. =6 Ministral 3 8B Instruct Q4_K_M8.5B total parameters75.0%
    3 tasks fewer passed
  8. #8 Gemma 4 E4B7.5B total parameters68.8%
    4 tasks fewer passed
  9. #9 Nemotron Nano 9B v2 Q4_08.9B total parameters37.5%
    9 tasks fewer passed
  10. #10 Qwen2.5 Coder 7B Instruct Q4_K_M7.6B total parameters0.0%
    15 tasks fewer passed

Other mode, this model only: thinking on · 100.0% · Task pass rate.

How this comparison works

Share of tasks that passed the recorded checks. Family comparisons show the labeled Instruct and reasoning versions. Other peer comparisons use the same requested thinking setting, not necessarily each model’s best mode. Small differences on these short tests may not hold on other tasks. Peer membership varies by test coverage and compatible settings.

Comparison saved with the 2026-10-12 assessment. Task-specific results.

Speed and task success: Everyday coding tasksThinking off · Matching test-machine label and settings · Across model sizes

Shorter task time is better only when the result is useful. These times include failures, tools and grading. Token generation speed measures a different part of the work.

Compare timings and outcomes
ModelTypical task timeGeneration speedTasks passed
LFM2.5 1.2B Q4_01.4 seconds16 of 16 tasks timed296.5 tokens/s2 of 1616 of 16 scored
Phi-4 Mini Instruct 3.8B Q4_K_M3.9 seconds16 of 16 tasks timed102.8 tokens/s0 of 1616 of 16 scored
SmolLM3 3B Q4_K_M4.5 seconds16 of 16 tasks timed117.2 tokens/s3 of 1616 of 16 scored
Granite 4.0 Micro 3.4B Q4_05.6 seconds16 of 16 tasks timed107.8 tokens/s4 of 1616 of 16 scored
Qwen3.5 0.8B Q4_05.7 seconds16 of 16 tasks timed236.1 tokens/s6 of 1616 of 16 scored
Ling 3.0 Tiny Q4_05.8 seconds16 of 16 tasks timed122.4 tokens/s15 of 1616 of 16 scored
Qwen2.5 Coder 7B Instruct Q4_K_M6.8 seconds16 of 16 tasks timed66.5 tokens/s0 of 1616 of 16 scored
Ornith 1.5 9B Q4_K_MThis model6.9 seconds16 of 16 tasks timed54.4 tokens/s15 of 1616 of 16 scored
Mellum 2.1 12B-A2.5B Thinking7.5 seconds16 of 16 tasks timed135.9 tokens/s14 of 1616 of 16 scored
Ministral 3 8B Instruct Q4_K_M8.6 seconds16 of 16 tasks timed54.7 tokens/s12 of 1616 of 16 scored
Nemotron Nano 9B v2 Q4_08.9 seconds16 of 16 tasks timed52.9 tokens/s6 of 1616 of 16 scored
Granite 4.2 8B Q4_09.2 seconds16 of 16 tasks timed53.2 tokens/s14 of 1616 of 16 scored
Gemma 4 12B QAT Q4_013.3 seconds16 of 16 tasks timed35.2 tokens/s13 of 1616 of 16 scored

Results from this test setup, not a promise for your computer. Hardware, software, tokenizers and answer length can change the ordering. Historical runtime builds were not always captured. Comparison saved with the 2026-10-12 assessment.

Test machine: AMD Ryzen 5 5600X (6 cores / 12 threads) · AMD Radeon RX 6700 XT (12 GB VRAM) · 32 GB or 128 GB system RAM, never mixed in one timing group. Full test machine details

With thinking on, larger-task success exceeds most tested medium-sized dense peers. In the matched timing cohort, Gemma 4 12B QAT Q4_0 completes more tasks but takes more than three times as long typically. The nearest lower-scoring alternatives are either only slightly faster or substantially slower, making this a useful speed-quality compromise.

Evidence for this judgment
Larger practical tasks · Thinking on
Average score63.6/100 · Mean task credit
Test coverage11 of 11 tasks scored
Task outcomes7 full credit · 0 partial · 4 no credit
Tasks timed11 tasks
Typical time per task (median)47.1 seconds
Explore these test results →

View this model’s answers, code and tool activity →

Larger practical tasks
Thinking on. Medium · 5–under 20B · dense. Rank 4 of 9 in this recorded comparison.

Speed: Larger practical tasks
Thinking on. Matching test-machine label and execution settings; spans model sizes. Timings include unsuccessful tasks, so read them alongside quality.

Larger practical tasksThinking on · Medium · 5–under 20B · dense

Build small apps, organize information and investigate files or web pages.

Mean task credit · 0–100 points · higher is better
11 of 11 tasks scored for each model.

Gaps are relative to this model. Treat small gaps cautiously.

  1. =1 Gemma 4 12B11.9B total parameters90.9/100
    27.3 points higher
  2. =1 Qwen3.5 9B9.0B total parameters90.9/100
    27.3 points higher
  3. #3 Gemma 4 12B QAT Q4_011.9B total parameters81.8/100
    18.2 points higher
  4. #4 Ornith 1.5 9B Q4_K_M9.2B total parameters63.6/100
    This model
  5. =5 Gemma 4 E4B7.5B total parameters54.5/100
    9.1 points lower
  6. =5 Ministral 3 14B Reasoning 2512 Q4_K_M13.5B total parameters54.5/100
    9.1 points lower
  7. #7 Nemotron Nano 9B v2 Q4_08.9B total parameters36.4/100
    27.3 points lower
  8. #8 Granite 4.2 8B Q4_08.8B total parameters27.3/100
    36.4 points lower
  9. #9 Ministral 3 8B Reasoning 2512 Q4_K_M8.5B total parameters18.2/100
    45.5 points lower

Other mode, this model only: thinking off · 54.5/100 · Mean task credit.

How this comparison works

Average credit per task; partly correct results can earn partial credit. Family comparisons show the labeled Instruct and reasoning versions. Other peer comparisons use the same requested thinking setting, not necessarily each model’s best mode. Small differences on these short tests may not hold on other tasks. Peer membership varies by test coverage and compatible settings.

Comparison saved with the 2026-10-12 assessment. Task-specific results.

Speed and task success: Larger practical tasksThinking on · Matching test-machine label and settings · Across model sizes

Shorter task time is better only when the result is useful. These times include failures, tools and grading. Token generation speed measures a different part of the work.

Compare timings and outcomes
ModelTypical task timeGeneration speedAverage score
LFM2.5 1.2B Thinking Q4_K_M12.5 seconds11 of 11 tasks timed291.2 tokens/s0.0/10011 of 11 scored
SmolLM3 3B Q4_K_M29.7 seconds11 of 11 tasks timed103.7 tokens/s9.1/10011 of 11 scored
Ling 3.0 Tiny Q4_044.1 seconds11 of 11 tasks timed129.0 tokens/s45.5/10011 of 11 scored
Ornith 1.5 9B Q4_K_MThis model47.1 seconds11 of 11 tasks timed52.9 tokens/s63.6/10011 of 11 scored
Mellum 2.1 12B-A2.5B Thinking55.7 seconds11 of 11 tasks timed133.9 tokens/s27.3/10011 of 11 scored
Nemotron Nano 9B v2 Q4_061.6 seconds11 of 11 tasks timed53.0 tokens/s36.4/10011 of 11 scored
Ministral 3 8B Reasoning 2512 Q4_K_M90.8 seconds11 of 11 tasks timed46.6 tokens/s18.2/10011 of 11 scored
Ministral 3 14B Reasoning 2512 Q4_K_M120.9 seconds11 of 11 tasks timed31.9 tokens/s54.5/10011 of 11 scored
Gemma 4 12B QAT Q4_0153.4 seconds11 of 11 tasks timed35.0 tokens/s81.8/10011 of 11 scored
Granite 4.2 8B Q4_0303.7 seconds11 of 11 tasks timed43.8 tokens/s27.3/10011 of 11 scored

Results from this test setup, not a promise for your computer. Hardware, software, tokenizers and answer length can change the ordering. Historical runtime builds were not always captured. Comparison saved with the 2026-10-12 assessment.

Test machine: AMD Ryzen 5 5600X (6 cores / 12 threads) · AMD Radeon RX 6700 XT (12 GB VRAM) · 32 GB or 128 GB system RAM, never mixed in one timing group. Full test machine details

Worth knowing

Publication redactions, unavailable components, source-link-only paper exports and withheld project files limit readers’ inspection of some examples.

Evidence for this judgment

Source limitation
Some selected evidence has recorded publication redactions.

How to interpret these assessments and their limitations →

How do the weight formats compare?

Everyday coding tasks · Task pass rate · requested / displayed reasoning on. Size is the weight download, not RAM. Different runtimes and placement can also affect results.

Q4_K_M · 5.78 GBLocal · test machine with 32 GB RAM · vulkan · 99 GPU layers · Fully scoredThinking on16/16 passed16/16 scored
Q5_K_M · 6.64 GBLocal · test machine with 32 GB RAM · vulkan · 99 GPU layers · Fully scoredThinking on16/16 passed16/16 scored
Q6_K · 7.56 GBLocal · test machine with 32 GB RAM · vulkan · 99 GPU layers · Fully scoredThinking on16/16 passed16/16 scored
Q8_0 · 9.79 GBLocal · test machine with 32 GB RAM · vulkan · 99 GPU layers · Fully scoredThinking on16/16 passed16/16 scored

Qualified rows show the reason they are outside the rankings; they cannot establish superiority over a fully tested format. Open the setup to inspect context, runtime and controls.

Choosing a quant

Shared assessment of Q4_K_M, Q5_K_M, Q6_K, Q8_0 · 2026-10-22 UTC. The current results table can include newer formats.

For small coding tasks, Q4_K_M is a sensible starting point when download size matters. Q5_K_M, Q6_K and Q8_0 also did well here. With reasoning requested on, every format passed every task; requested off, Q6_K and Q8_0 led by only a single task. That narrow gap offers little reason to download a larger format for small edits alone, but this limited bank cannot settle performance on other tasks.

Evidence for this judgment

Everyday coding tasks · Thinking requested on

Matching task controls and 16,384 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth.

Weight formatTask pass rateTask timeDownloadInspect
Q4_K_MSelected formatornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
16/16 passed16/16 scored13.3s16/16 timed5.78 GBWeights, not required RAMAnswers and code
Q5_K_Mornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
16/16 passed16/16 scored12.6s16/16 timed6.64 GBWeights, not required RAMAnswers and code
Q6_Kornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
16/16 passed16/16 scored16.2s16/16 timed7.56 GBWeights, not required RAMAnswers and code
Q8_0ornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
16/16 passed16/16 scored27.2s16/16 timed9.79 GBWeights, not required RAMAnswers and code

Everyday coding tasks · Thinking requested off

Matching task controls and 16,384 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth.

Weight formatTask pass rateTask timeDownloadInspect
Q4_K_MSelected formatornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
15/16 passed16/16 scored6.9s16/16 timed5.78 GBWeights, not required RAMAnswers and code
Q5_K_Mornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
15/16 passed16/16 scored10.0s16/16 timed6.64 GBWeights, not required RAMAnswers and code
Q6_Kornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
16/16 passed16/16 scored8.8s16/16 timed7.56 GBWeights, not required RAMAnswers and code
Q8_0ornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
16/16 passed16/16 scored12.7s16/16 timed9.79 GBWeights, not required RAMAnswers and code

For larger tool-driven tasks with reasoning requested off, favor Q5_K_M. It scored appreciably higher than the other formats, with a smaller download and shorter median task time than Q6_K or Q8_0.

Evidence for this judgment

Larger practical tasks · Thinking requested off

Matching task controls and 16,384 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth.

Weight formatMean task creditTask timeDownloadInspect
Q4_K_MSelected formatornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
54.5/10011/11 scored32.2s11/11 timed5.78 GBWeights, not required RAMAnswers and code
Q5_K_Mornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
81.8/10011/11 scored38.8s11/11 timed6.64 GBWeights, not required RAMAnswers and code
Q6_Kornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
45.5/10011/11 scored50.9s11/11 timed7.56 GBWeights, not required RAMAnswers and code
Q8_0ornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
54.5/10011/11 scored46.9s11/11 timed9.79 GBWeights, not required RAMAnswers and code

With reasoning requested on, Q8_0 led larger tool-driven tasks, making its larger download worth considering for that use. Q6_K trailed by only a single task; Q4_K_M and Q5_K_M tied below them. This supports considering Q8_0 for this task set, not treating it as a universal upgrade.

Evidence for this judgment

Larger practical tasks · Thinking requested on

Matching task controls and 16,384 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth.

Weight formatMean task creditTask timeDownloadInspect
Q4_K_MSelected formatornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
63.6/10011/11 scored47.1s11/11 timed5.78 GBWeights, not required RAMAnswers and code
Q5_K_Mornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
63.6/10011/11 scored66.0s11/11 timed6.64 GBWeights, not required RAMAnswers and code
Q6_Kornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
72.7/10011/11 scored65.7s11/11 timed7.56 GBWeights, not required RAMAnswers and code
Q8_0ornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
81.8/10011/11 scored58.7s11/11 timed9.79 GBWeights, not required RAMAnswers and code

For logic and number puzzles with reasoning requested on, Q5_K_M offers an attractive tradeoff: it scored highest and had shorter median task time than Q6_K and Q8_0. Compared with Q4_K_M, its higher score came with a larger download and longer median task time.

Evidence for this judgment

Logic and number puzzles · Thinking requested on

Matching task controls and 16,384 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth.

Weight formatMean puzzle creditTask timeDownloadInspect
Q4_K_MSelected formatornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
73.5/10034/34 scored40.6s34/34 timed5.78 GBWeights, not required RAMAnswers and code
Q5_K_Mornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
85.3/10034/34 scored47.4s34/34 timed6.64 GBWeights, not required RAMAnswers and code
Q6_Kornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
76.5/10034/34 scored50.4s34/34 timed7.56 GBWeights, not required RAMAnswers and code
Q8_0ornith-aivulkan · 32 GB test machineCPU MoE layers: Not recorded
Recorded execution settings
Context
16,384
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
79.4/10034/34 scored69.7s34/34 timed9.79 GBWeights, not required RAMAnswers and code
Scope and limits
  • These are comparisons of tested setups registered to one checkpoint. They do not isolate bit depth from conversion, publisher, runtime-build or template differences.
  • Only complete eligible banks with matching task controls and recorded context enter the quality comparison. Execution profiles are shown per setup; timing is compared only when every recorded execution profile matches and timing coverage is complete. Other results remain coverage, not ranked evidence.
  • Download size is not RAM or VRAM requirement. No memory-fit recommendation is established by this comparison.
  • Small score differences can reflect a single question. Similar results are not proof of equivalence; timings describe this benchmark machine.
About this quant assessment
Requested model
gpt-6-astra
Returned model
Not supplied; request identity only.
Exact provider snapshot
Not supplied.
Generation route
Codex CLI · ChatGPT account
Prompt version
quant-summary-v3
Review time
2026-10-22T06:58:37.283045+00:00
Evidence reviewer
Codex (agent)
Publication authorization
User · 2026-10-22T05:55:43.625712+00:00 · Go ahead and complete the remaining work. Spin up agents: luna, terra, sol, astra, as needed to get it done. Narrative summaries: these are the astra generated summaries? Go ahead and get those completed as well. Once all of this is done, you may push the completed result.
Source edition
dataset-5a6f96ec56c3d240cc42fa23147c1495da1be71f3ef0722c0e5f8c4f5e126a83
Revision
quant-summary-ba200a7c5ca49f0959bdc8bc6baabb9bf91995e376dddf34a1a3c50f5283de2e

Thinking tradeoff

16/16 matched scored questions · 16/16 matched timed questions

Thinking offMean matched task credit93.8%
Thinking onMean matched task credit100.0%

Thinking on took 9.0 seconds more per matched timed question, on average.

All matched quality and timing measurements
Matched measurementRecorded result
Average credit, off → on93.8% → 100.0%
Mean paired credit change+6.3 percentage points
Question outcomes1 improved · 15 unchanged · 0 worse
Mean added task time+9.0s
Mean task time, off → on10.4s → 19.4s
Median task time, off → on6.9s → 13.3s
Median of paired time changes+5.2s
Total added task time+143.7s
Ratio of summed times, on / off1.9×

Quality and timing can cover different subsets. Median paired change is calculated per question; it is not the difference between the two medians. Task time includes tools and grading.

Inspect the matched modes and their answers →
Inspect individual matched questions

Weight formats compared

All tested weight formats of this checkpoint, using the selected thinking setting. This shared comparison updates with new results; each saved model summary keeps its original date.

Everyday coding tasks

Scores and times describe this task bank. Quick failures are not a speed advantage. Different execution settings are labeled.

VersionTask resultTypical task timeWeight downloadExplore
Ornith 1.5 9B Q5_K_MThinking on100.0%16/16 scored12.6 seconds16/16 timed6.64 GBAnswers and code

Compare

Ornith 1.5 9B Q6_KThinking on100.0%16/16 scored16.2 seconds16/16 timed7.56 GBAnswers and code

Compare

Ornith 1.5 9B Q8_0Thinking on100.0%16/16 scored27.2 seconds16/16 timed9.79 GBAnswers and code

Compare

Ornith 1.5 9B Q4_K_MThinking on100.0%16/16 scored13.3 seconds16/16 timed5.78 GBAnswers and code

Sampled repetitions

Repeated runs are a separate experiment and never add votes to the baseline score.

Thinking on · temperature 0.6

48/48 repeats scored · 47 full-credit repeats.

16/16 questions have every planned repeat scored. Of those, 15 earned full credit on every repeat.

Recorded repetition conditions
  • Seed 303 · requested on · verified
  • Seed 101 · requested on · verified
  • Seed 202 · requested on · verified

Everyday coding tasks: recorded outcomes

Find information in a small project, edit files and fix bugs.

16 full credit

Each segment counts tasks, not credit points. Partial credit varies by task.

Scoring and fixed question bank →

This model’s answers, code and tool activity

Select a question to read the original answer, tool calls and available code changes for this configuration and thinking setting.

1–16 of 16 matching questions · 16 in this version

  • Locate the Retry Limit

    The user asks where the project's retry-limit setting is defined and what value it holds. This exercises locating a named setting in code and reporting both its location and value.

    basic-T1 · primary
  • Find the Release Date

    The user asks for the release date mentioned in a project notes document. This exercises reading a specified file and extracting a particular fact rather than guessing from outside knowledge.

    basic-T2 · primary
  • Change the Retry Limit

    The user asks to change the project's retry limit to a specified value. This exercises finding the relevant setting and making a focused code edit rather than merely describing the change.

    basic-T3 · primary
  • Create an Exact One-Line File

    The user asks to create a text file containing exactly one supplied line. This exercises basic file creation and precise instruction following, including avoiding extra text in the file.

    basic-T4 · primary
  • Explain the Latest Commit

    The user asks what the most recent Git commit changed in the project. This exercises inspecting recorded changes and turning them into a clear explanation of what was modified.

    basic-T5 · primary
  • Count and Compare Python Files

    The user asks how many Python files are in the application folder and which is largest. This exercises identifying relevant files, counting them, and comparing their sizes.

    basic-T6 · primary
  • Add Readable Byte Sizes

    The user asks to add a function that converts byte counts into readable sizes and remove the related TODO comment. This exercises implementing a small utility and updating the surrounding code consistently.

    basic-T7 · primary
  • Collect TODO Comments

    The user asks to find every TODO comment in the Python code and append each to the project's task list on a separate line. This combines searching across files with collecting text into an existing document.

    basic-T8 · primary
  • Fix Code Behind a Failing Test

    The user asks to investigate a failing test and fix the application code without changing the test. This exercises diagnosing a bug while respecting an explicit limit on what may be edited.

    basic-H1 · primary
  • Rename a Function Throughout

    The user asks to rename a network function and update every place in the project that uses it. This exercises tracking references across files and making a coordinated change rather than editing only the definition.

    basic-H2 · primary
  • Find the Proxy Settings

    The user asks to find a proxy configuration file somewhere under the project's parent folder and report its host and port. This exercises searching beyond the project folder and reading configuration values.

    basic-H3 · primary
  • Locate an Error in a Log

    The user asks which error code appears in a long event log and on which line. This exercises finding a specific entry in a large file and reporting its precise location.

    basic-H4 · primary
  • Append a Final Task

    The user asks to append a supplied release-related task as the final line of an existing task list. This exercises preserving existing content while adding exact text in the requested position.

    basic-H5 · primary
  • Calculate Total Retry Waiting

    The user asks for the total time a network function spends sleeping when all attempts fail, worked out from the code. This exercises following retry logic and adding the delays along that failure path.

    basic-H6 · primary
  • Document Every Command-Line Option

    The user asks for a project readme listing every command-line option with its help text. This exercises discovering the program's available options and transferring that information into complete, source-grounded documentation.

    basic-H7 · primary
  • Undo a Commit by Editing

    The user asks to undo the latest Git commit's change by editing the file directly, without using Git to make changes. This exercises understanding a previous edit and reversing it under a method constraint.

    basic-H8 · primary
Filters change this list, not the fixed scoring denominator. Search currently covers question IDs, titles and available task explanations. Focus a control to load prompt/outcome search where this edition provides it.

Suggested comparisons

These are starting points for inspection, not claims that one setup is better. Runtime and placement differences remain visible.

Thinking on versus off

Verified requested on/off conditions of the same configuration, matched on the recorded test protocol.

Requested thinking off · Complete compared coverage

Compare these configurations →

Selection rule and evidence IDsPolicy comparisons-v3 · Template verified-thinking-pair
Facts: paired-9940f7caf067465f44368404
Matching candidate set: paired-9940f7caf067465f44368404

Ornith 1.5 9B Q5_K_M

Another tested weight format of this checkpoint. Complete results first, then nearest download size.

Requested thinking on · Complete compared coverage

Compare these configurations →

Selection rule and evidence IDsPolicy comparisons-v3 · Template same-checkpoint-matching-controls
Facts: fact-8911d11609395c5679b85df7, fact-328e4a9121e69c97cd7319be
Matching candidate set: fact-328e4a9121e69c97cd7319be, fact-81d2c9b418915e652817cddd, fact-c7309f96857ab4d57c685d63

Qwen3.5 9B UD-Q4_K_XL

Another checkpoint in the same total-parameter band and architecture category, with matching task controls.

Requested thinking on · Complete compared coverage

Compare these configurations →

Selection rule and evidence IDsPolicy comparisons-v3 · Template same-parameter-architecture-matching-controls
Facts: fact-8911d11609395c5679b85df7, fact-f7b0f0121a8f96c9dfdb8666
Matching candidate set: fact-f7b0f0121a8f96c9dfdb8666, fact-470bad2baa7a5b9251437b89, fact-6650b707149de7ed1c733b20, fact-10a001757ce5e3a4ad0b8662, fact-291755b68033af9343a74237, fact-4afd936ddf9e3ca0a6f4a16c, fact-e28b7d3cf40620276ac6c151, fact-d2b59d44b662ae713aeac31a, fact-d91472a9baf4c9f2b588d3e7, fact-d0f093eba79679ef590e1554, fact-4dd123d36ba8eb2eb1ba71b3, fact-cec7a0963364769cc87c775c, fact-503ad2bbd095726e4510f2a5, fact-32c8729370f3fa2ea4742bf7, fact-97b75588d5514b8728fe4aa3, fact-300ea84f20ecc412e99cd12e, fact-167aad313fe901c497cd751b

Tested weights

Q4_K_M · 5.78 GB

ornith-ai · recorded source

Ornith-1.5-9B-Q4_K_M.gguf · Exact GGUF file →Historical first-MiB hash only; not a full-file identity.

Q5_K_M · 6.64 GB

ornith-ai · recorded source

Ornith-1.5-9B-Q5_K_M.gguf · Exact GGUF file →Historical first-MiB hash only; not a full-file identity.

Q6_K · 7.56 GB

ornith-ai · recorded source

Ornith-1.5-9B-Q6_K.gguf · Exact GGUF file →Historical first-MiB hash only; not a full-file identity.

Q8_0 · 9.79 GB

ornith-ai · recorded source

Ornith-1.5-9B-Q8_0.gguf · Exact GGUF file →Historical first-MiB hash only; not a full-file identity.

Explore the answers

Choose a suite to inspect its exact available questions, references and model answers.