AgentBench / results
← All models

Gemma 4 19B-A4B REAP

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

Choose a tested configuration

Get this tested model

Q4_K_S · 11.34 GB download

Links follow the selected quant and configuration.

Setup, memory & provenance · Q4_K_S

Q4_K_S · 11.34 GB of weights · rocm · context 32768 · GPU layers 99

Medium · 5–under 20B · MoE · 18.4B total · Active parameter count not recorded

Memory fit has not been verified for this configuration.

partial

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: inconclusive · probe 2 · 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

18.1 / 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 offDetails
Everyday coding tasksTask pass rate
Pass rate
14/16 · 88%
Answers completed
15/16 · 94%
Correct / completed
14/15 · 93%
14 full credit1 incorrect1 failed to finish
16/16 scored8.8s median task time · 16 measured
Larger practical tasksMean task credit
Average credit
45.5%
Full-credit rate
5/11 · 45%
Answers completed
11/11 · 100%
Full credit / completed
5/11 · 45%
5 full credit6 incorrect
11/11 scored37.0s median task time · 11 measured
Logic and number puzzlesMean puzzle credit
Average credit
0.2%
Full-credit rate
0/34 · 0%
Answers completed
34/34 · 100%
Full credit / completed
0/34 · 0%
6 partial credit28 incorrect
34/34 scored102.2s median task time · 34 measured
Published hard problemsMean answer credit
Average credit
0.3%
Full-credit rate
0/11 · 0%
Answers completed
11/11 · 100%
Full credit / completed
0/11 · 0%
1 partial credit10 incorrect
11/11 scored102.3s median task time · 11 measured
Following exact instructions · original campaignStrict prompt compliance
Pass rate
1/23 · 4%
Answers completed
23/23 · 100%
Correct / completed
1/23 · 4%
1 full credit22 incorrect
23/23 scored101.0s median task time · 23 measured
Reading without guessingExact-match accuracy
Pass rate
0/16 · 0%
Answers completed
16/16 · 100%
Correct / completed
0/16 · 0%
16 incorrect
16/16 scored100.9s median task time · 16 measured
Calculating from a passageNumeric exact-match accuracy
Pass rate
0/7 · 0%
Answers completed
7/7 · 100%
Correct / completed
0/7 · 0%
7 incorrect
7/7 scored103.1s median task time · 7 measured
Reading research papersAnswer token F1
Average credit
0.0%
Full-credit rate
0/8 · 0%
Answers completed
8/8 · 100%
Full credit / completed
0/8 · 0%
8 incorrect
8/8 scored105.7s median task time · 8 measured
Resisting incorrect correctionsAll-branches pass rate
Pass rate
3/12 · 25%
Answers completed
12/12 · 100%
Correct / completed
3/12 · 25%
3 full credit9 incorrect
12/12 scored2.0s 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.

Consider this Q4_K_S setup for small coding jobs with human checking, not as a general-purpose assistant. With thinking requested off, it passed most everyday coding tasks, with a median total task time of 8.8 seconds on this test machine—a useful balance for bounded work. Larger practical tasks succeeded less than half the time. Passage reading and puzzles were extremely poor, sharply limiting its usefulness beyond coding.

Evidence for this judgment
Everyday coding tasks · Thinking off
Tasks passed14 of 16 (87.5%)
Test coverage16 of 16 tasks scored
Tasks timed16 tasks
Typical time per task (median)8.8 seconds
Explore these test results →

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

Larger practical tasks · Thinking off
Average score45.5/100 · Mean task credit
Test coverage11 of 11 tasks scored
Task outcomes5 full credit · 0 partial · 6 no credit
Explore these test results →

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

Reading without guessing · Thinking off
Tasks passed0 of 16 (0.0%)
Test coverage16 of 16 tasks scored
Explore these test results →

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

Logic and number puzzles · Thinking off
Average score0.2/100 · Mean puzzle credit
Test coverage34 of 34 tasks scored
Task outcomes0 full credit · 6 partial · 28 no credit
Explore these test results →

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

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

About this summary

The original analyst response failed condition-reference validation. Codex removed references to other models’ conditions offline, preserving the prose and measurements. The failed job and original receipts remain recorded; no additional model request was made.

Analyst provider
openai
Generation route
Codex CLI · codex-cli 0.163.0 · 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-v8
Generation time
2026-10-22T05:45:13.366235Z
Review time
2026-10-22T05:55:44.057681+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-5a6f96ec56c3d240cc42fa23147c1495da1be71f3ef0722c0e5f8c4f5e126a83
Analysis input hash
7642e172df96a735e737642e29c3dd961c86e5fa3b75f941752cb20e9fc07a51
Summary revision
summary-fda0a914bc39235a77ebb86f223b511ad298bba2a05ffb9a8a9a7a98b289b4cb
Original response status
Completed response · failed reference validation
Original job
job-128182742001d5d3cbad28f0779d25a7b681a14f9e9d60ff7bc243167d182aad
Original receipt hash
eaa97894824607a4920058f1b73b4437c7429da2dc3413a0b6fcb4213fd71edb
Original content hash
79def12d820eec5a7e73a5f11dd9af5e71a46019f5244bb4e230857f9e89239e
Offline editorial change
Codex · foreign condition references only · peer_comparisons[0], peer_comparisons[1]

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

Good at

Useful for bounded coding assistance: it passed most small-project tasks involving finding information, editing files and fixing bugs. That supports supervised coding help, rather than assuming it can manage an entire project.

Evidence for this judgment
Everyday coding tasks · Thinking off
Tasks passed14 of 16 (87.5%)
Test coverage16 of 16 tasks scored
Incorrect completed answers1 tasks
Failed to finish1 tasks
Explore these test results →

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

Struggles with

Do not rely on this setup for passage answers, calculations from text, research-paper questions, or difficult math and puzzles. Scores were zero or near zero across these tests, not merely uneven.

Evidence for this judgment
Reading without guessing · Thinking off
Tasks passed0 of 16 (0.0%)
Test coverage16 of 16 tasks scored
Explore these test results →

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

Calculating from a passage · Thinking off
Tasks passed0 of 7 (0.0%)
Test coverage7 of 7 tasks scored
Explore these test results →

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

Reading research papers · Thinking off
Average score0.0/100 · Answer token F1
Test coverage8 of 8 tasks scored
Task outcomes0 full credit · 0 partial · 8 no credit
Explore these test results →

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

Logic and number puzzles · Thinking off
Average score0.2/100 · Mean puzzle credit
Test coverage34 of 34 tasks scored
Task outcomes0 full credit · 6 partial · 28 no credit
Explore these test results →

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

Published hard problems · Thinking off
Average score0.3/100 · Mean answer credit
Test coverage11 of 11 tasks scored
Task outcomes0 full credit · 1 partial · 10 no credit
Explore these test results →

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

Strict instructions and source-backed corrections are substantial weaknesses. It rarely met every wording or formatting requirement, and usually failed the combined test of resisting false corrections while updating answers when the source changed.

Evidence for this judgment
Following exact instructions · Thinking off
Tasks passed1 of 23 (4.3%)
Test coverage23 of 23 tasks scored
Explore these test results →

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

Resisting incorrect corrections · Thinking off
Tasks passed3 of 12 (25.0%)
Test coverage12 of 12 tasks scored
Explore these test results →

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

Compared with other models

With thinking requested off, small-coding results were close to the other tested medium-sized MoE models. Its larger-task score was moderately higher, but that advantage on a small test does not establish broad superiority.

Evidence for this judgment

Everyday coding tasks
Thinking off. Medium · 5–under 20B · MoE (total parameters). Tied at rank 2 of 3 in this recorded comparison.

Larger practical tasks
Thinking off. Medium · 5–under 20B · MoE (total parameters). Rank 1 of 3 in this recorded comparison.

Everyday coding tasks · Thinking off
Tasks passed14 of 16 (87.5%)
Test coverage16 of 16 tasks scored
Explore these test results →

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

Larger practical tasks · Thinking off
Average score45.5/100 · Mean task credit
Test coverage11 of 11 tasks scored
Task outcomes5 full credit · 0 partial · 6 no credit
Explore these test results →

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

Everyday coding tasksThinking off · Medium · 5–under 20B · MoE (total parameters)

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 Ling 3.0 Tiny Q4_07.9B total parameters93.8%
    1 task more passed
  2. =2 Gemma 4 19B-A4B REAP (Q4_K_S)18.4B total parameters87.5%
    This model
  3. =2 Mellum 2.1 12B-A2.5B Thinking12.1B total parameters87.5%
    Same score
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.

MoE models use only part of their parameters for each token. This group uses total size; active counts are not included in this saved comparison.

Comparison saved with the 2026-10-22 assessment. Task-specific results.

Larger practical tasksThinking off · Medium · 5–under 20B · MoE (total parameters)

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 19B-A4B REAP (Q4_K_S)18.4B total parameters45.5/100
    This model
  2. =2 Ling 3.0 Tiny Q4_07.9B total parameters27.3/100
    18.2 points lower
  3. =2 Mellum 2.1 12B-A2.5B Thinking12.1B total parameters27.3/100
    18.2 points lower
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.

MoE models use only part of their parameters for each token. This group uses total size; active counts are not included in this saved comparison.

Comparison saved with the 2026-10-22 assessment. Task-specific results.

For passage reading and puzzles, this requested-off setup scored far below the other tested medium-sized MoE models. The larger-task advantage is therefore a narrow reason to choose it, not a general capability advantage.

Evidence for this judgment

Reading without guessing
Thinking off. Medium · 5–under 20B · MoE (total parameters). Rank 3 of 3 in this recorded comparison.

Logic and number puzzles
Thinking off. Medium · 5–under 20B · MoE (total parameters). Rank 3 of 3 in this recorded comparison.

Reading without guessing · Thinking off
Tasks passed0 of 16 (0.0%)
Test coverage16 of 16 tasks scored
Explore these test results →

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

Logic and number puzzles · Thinking off
Average score0.2/100 · Mean puzzle credit
Test coverage34 of 34 tasks scored
Task outcomes0 full credit · 6 partial · 28 no credit
Explore these test results →

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

Reading without guessingThinking off · Medium · 5–under 20B · MoE (total parameters)

Answer from a passage—or say that the passage does not contain the answer.

Exact-match accuracy · 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 Mellum 2.1 12B-A2.5B Thinking12.1B total parameters81.3%
    13 tasks more passed
  2. #2 Ling 3.0 Tiny Q4_07.9B total parameters62.5%
    10 tasks more passed
  3. #3 Gemma 4 19B-A4B REAP (Q4_K_S)18.4B total parameters0.0%
    This model
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.

MoE models use only part of their parameters for each token. This group uses total size; active counts are not included in this saved comparison.

Comparison saved with the 2026-10-22 assessment. Task-specific results.

Logic and number puzzlesThinking off · Medium · 5–under 20B · MoE (total parameters)

Solve newly generated puzzles using arithmetic, deduction and planning.

Mean puzzle credit · 0–100 points · higher is better
34 of 34 tasks scored for each model.

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

  1. #1 Mellum 2.1 12B-A2.5B Thinking12.1B total parameters41.2/100
    41.0 points higher
  2. #2 Ling 3.0 Tiny Q4_07.9B total parameters29.6/100
    29.4 points higher
  3. #3 Gemma 4 19B-A4B REAP (Q4_K_S)18.4B total parameters0.2/100
    This model
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.

MoE models use only part of their parameters for each token. This group uses total size; active counts are not included in this saved comparison.

Comparison saved with the 2026-10-22 assessment. Task-specific results.

Worth knowing

The thinking comparison remains an untested preview, so enabling thinking cannot be recommended from these results. Reasoning control is unverified, and local output-protocol incompatibility does not establish defective weights or insufficient memory.

Evidence for this judgment

Source limitation
Some suite/condition results are incomplete or ineligible; their observed scores are previews.

Source limitation
Some requested reasoning conditions lack verified control.

Source limitation
Local output-protocol incompatibility is classified as a model/configuration failure. Thinking comparison remains untested. Technical failures stay unscored; this is not evidence of defective weights or out-of-memory failure. No additional parser rerolls are authorized.

Publication redactions limit readers’ inspection of some underlying evidence.

Evidence for this judgment

Source limitation
Some selected evidence has recorded publication redactions.

Context-exhaustion cases allow a same-budget retry. Continued failure to finish is distinct from an incorrect answer; this follow-up was selected after failures.

Evidence for this judgment

Source limitation
One same-budget retry after context exhaustion. A second context limit is failed to answer (zero task-completion credit), not an incorrect submitted answer. Answer-only averages exclude these no-answer outcomes; completed retry answers enter the normal scoring path. This follow-up is selected after observing failures.

How to interpret these assessments and their limitations →

How do the weight formats compare?

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

IQ4_XS · 10.23 GBLocal · test machine with 32 GB RAM · rocm · 99 GPU layers · Recorded · some results unscored or qualifiedThinking off12/15 passed15/16 scored · 1 technical / answer-channel errors
Q4_K_S · 11.34 GBLocal · test machine with 32 GB RAM · rocm · 99 GPU layers · Recorded · some results unscored or qualifiedThinking off14/16 passed16/16 scored
Q5_K_M · 13.99 GBLocal · test machine with 32 GB RAM · rocm · 99 GPU layers · Recorded · some results unscored or qualifiedThinking off15/16 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 IQ4_XS, Q4_K_S, Q5_K_M · 2026-10-22 UTC. The current results table can include newer formats.

For these requested-off setups, choose by task rather than bit depth. Q4_K_S and Q5_K_M are reasonable options for small coding jobs; IQ4_XS offers a smaller-download starting point for supervised larger tool-driven work. The larger Q5_K_M download is more compelling for the selected hard math and reasoning problems, where it clearly led, than for larger practical tasks. Reasoning suppression was not verified; the thinking comparison remains untested.

Evidence for this judgment

Everyday coding tasks · Thinking requested off

Matching task controls and 32,768 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth. Task times are not comparable for this group.

Weight formatTask pass rateTask timeDownloadInspect
Q4_K_SSelected formatmradermacherrocm · 32 GB test machineCPU MoE layers: 4
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
14/16 passed16/16 scoredNot compared16/16 timed11.34 GBWeights, not required RAMAnswers and code
Q5_K_Mmradermacherrocm · 32 GB test machineCPU MoE layers: 11
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
15/16 passed16/16 scoredNot compared16/16 timed13.99 GBWeights, not required RAMAnswers and code

Larger practical tasks · Thinking requested off

Matching task controls and 32,768 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth. Task times are not comparable for this group.

Weight formatMean task creditTask timeDownloadInspect
IQ4_XSmradermacherrocm · 32 GB test machineCPU MoE layers: 1
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
45.5/10011/11 scoredNot compared11/11 timed10.23 GBWeights, not required RAMAnswers and code
Q4_K_SSelected formatmradermacherrocm · 32 GB test machineCPU MoE layers: 4
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
45.5/10011/11 scoredNot compared11/11 timed11.34 GBWeights, not required RAMAnswers and code
Q5_K_Mmradermacherrocm · 32 GB test machineCPU MoE layers: 11
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
36.4/10011/11 scoredNot compared11/11 timed13.99 GBWeights, not required RAMAnswers and code

Published hard problems · Thinking requested off

Matching task controls and 32,768 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth. Task times are not comparable for this group.

Weight formatMean answer creditTask timeDownloadInspect
IQ4_XSmradermacherrocm · 32 GB test machineCPU MoE layers: 1
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
30.7/10011/11 scoredNot compared11/11 timed10.23 GBWeights, not required RAMAnswers and code
Q4_K_SSelected formatmradermacherrocm · 32 GB test machineCPU MoE layers: 4
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
0.3/10011/11 scoredNot compared11/11 timed11.34 GBWeights, not required RAMAnswers and code
Q5_K_Mmradermacherrocm · 32 GB test machineCPU MoE layers: 11
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
54.5/10011/11 scoredNot compared11/11 timed13.99 GBWeights, not required RAMAnswers and code

For small coding jobs, Q4_K_S offers a smaller download than Q5_K_M, whose lead amounts to a single question. That narrow advantage is not a compelling upgrade by itself.

Evidence for this judgment

Everyday coding tasks · Thinking requested off

Matching task controls and 32,768 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth. Task times are not comparable for this group.

Weight formatTask pass rateTask timeDownloadInspect
Q4_K_SSelected formatmradermacherrocm · 32 GB test machineCPU MoE layers: 4
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
14/16 passed16/16 scoredNot compared16/16 timed11.34 GBWeights, not required RAMAnswers and code
Q5_K_Mmradermacherrocm · 32 GB test machineCPU MoE layers: 11
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
15/16 passed16/16 scoredNot compared16/16 timed13.99 GBWeights, not required RAMAnswers and code

For larger tool-driven tasks, IQ4_XS and Q4_K_S earned matching aggregate credit, with Q5_K_M narrowly behind. IQ4_XS is the smallest download and a reasonable starting point here, but the limited task success warrants supervision with any format.

Evidence for this judgment

Larger practical tasks · Thinking requested off

Matching task controls and 32,768 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth. Task times are not comparable for this group.

Weight formatMean task creditTask timeDownloadInspect
IQ4_XSmradermacherrocm · 32 GB test machineCPU MoE layers: 1
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
45.5/10011/11 scoredNot compared11/11 timed10.23 GBWeights, not required RAMAnswers and code
Q4_K_SSelected formatmradermacherrocm · 32 GB test machineCPU MoE layers: 4
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
45.5/10011/11 scoredNot compared11/11 timed11.34 GBWeights, not required RAMAnswers and code
Q5_K_Mmradermacherrocm · 32 GB test machineCPU MoE layers: 11
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
36.4/10011/11 scoredNot compared11/11 timed13.99 GBWeights, not required RAMAnswers and code

Prefer IQ4_XS or Q5_K_M over Q4_K_S for passage answering and exact instruction-following: their tested setups scored substantially higher. IQ4_XS also led Q5_K_M on instruction-following despite its smaller download.

Evidence for this judgment

Reading without guessing · Thinking requested off

Matching task controls and 32,768 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth. Task times are not comparable for this group.

Weight formatExact-match accuracyTask timeDownloadInspect
IQ4_XSmradermacherrocm · 32 GB test machineCPU MoE layers: 1
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
13/16 passed16/16 scoredNot compared16/16 timed10.23 GBWeights, not required RAMAnswers and code
Q4_K_SSelected formatmradermacherrocm · 32 GB test machineCPU MoE layers: 4
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
0/16 passed16/16 scoredNot compared16/16 timed11.34 GBWeights, not required RAMAnswers and code
Q5_K_Mmradermacherrocm · 32 GB test machineCPU MoE layers: 11
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
14/16 passed16/16 scoredNot compared16/16 timed13.99 GBWeights, not required RAMAnswers and code

Following exact instructions · Thinking requested off

Matching task controls and 32,768 context. These quality results describe the recorded setups; they do not isolate the effect of bit depth. Task times are not comparable for this group.

Weight formatStrict prompt complianceTask timeDownloadInspect
IQ4_XSmradermacherrocm · 32 GB test machineCPU MoE layers: 1
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
15/23 passed23/23 scoredNot compared23/23 timed10.23 GBWeights, not required RAMAnswers and code
Q4_K_SSelected formatmradermacherrocm · 32 GB test machineCPU MoE layers: 4
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
1/23 passed23/23 scoredNot compared23/23 timed11.34 GBWeights, not required RAMAnswers and code
Q5_K_Mmradermacherrocm · 32 GB test machineCPU MoE layers: 11
Recorded execution settings
Context
32,768
GPU layer setting
99
Threads
6
Concurrency
1
Runtime build
Not recorded
Tool adapter
saved-campaign-agent-tools
12/23 passed23/23 scoredNot compared23/23 timed13.99 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
2026-10-22T06:58:37.390240+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-99f0427570aaf7fe2d9b196524cf540c0b4be86f6b8138c6c01c3bad05e0f2e4

What changed in these results

One same-budget retry after context exhaustion. A second context limit is failed to answer (zero task-completion credit), not an incorrect submitted answer. Answer-only averages exclude these no-answer outcomes; completed retry answers enter the normal scoring path. This follow-up is selected after observing failures.

Local output-protocol incompatibility is classified as a model/configuration failure. Thinking comparison remains untested. Technical failures stay unscored; this is not evidence of defective weights or out-of-memory failure. No additional parser rerolls are authorized.

Thinking tradeoff

No verified on/off comparison is available for this configuration and test suite. Requested modes alone do not establish a controlled thinking comparison.

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
Gemma 4 19B-A4B REAP (Q4_K_S)Thinking off87.5%16/16 scored8.8 seconds16/16 timed11.34 GBAnswers and code
Gemma 4 19B-A4B REAP (Q5_K_M)Thinking off · Different or uncertain execution settings93.8%16/16 scored13.8 seconds16/16 timed13.99 GBAnswers and code

Compare

Gemma 4 19B-A4B REAPThinking off · Different or uncertain execution settings80.0%15/16 scored · Incomplete / not ranked6.3 seconds15/16 timed10.23 GBAnswers and code

Compare

Sampled repetitions

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

Thinking off · temperature 0.6

48/48 repeats scored · 44 full-credit repeats.

16/16 questions have every planned repeat scored. Of those, 12 earned full credit on every repeat.

Recorded repetition conditions
  • Seed 202 · requested off · inconclusive
  • Seed 303 · requested off · inconclusive
  • Seed 101 · requested off · inconclusive

Everyday coding tasks: recorded outcomes

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

14 full credit1 incorrect1 failed to finish

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.

Gemma 4 19B-A4B REAP (Q5_K_M)

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

Requested thinking off · Complete compared coverage

Compare these configurations →

Selection rule and evidence IDsPolicy comparisons-v3 · Template same-checkpoint-matching-controls
Facts: fact-69b7b20fcb8961feae28e4a8, fact-57f1be0ca8bc0e7cb0e343b4
Matching candidate set: fact-57f1be0ca8bc0e7cb0e343b4, fact-45e16a9e4c4c3e1b61f0f139

Mellum 2.1 12B-A2.5B Thinking Q4_K_M

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

Requested thinking off · Complete compared coverage

Compare these configurations →

Selection rule and evidence IDsPolicy comparisons-v3 · Template same-parameter-architecture-matching-controls
Facts: fact-69b7b20fcb8961feae28e4a8, fact-f2afff5be73b3b2bf0133034
Matching candidate set: fact-f2afff5be73b3b2bf0133034, fact-18b9d27661127ac14e5e16ab, fact-1ec245f5a9717680a5818e83, fact-48a3c8ec09a4a055e278d563

Tested weights

IQ4_XS · 10.23 GB

mradermacher · recorded source

657a339b7f44701d3491f148dbadd6f6440f45e8fba96a5eda21428bc76b4c65 · Exact GGUF file →Historical first-MiB hash only; not a full-file identity.

Q4_K_S · 11.34 GB

mradermacher · recorded source

gemma-4-19b-a4b-it-REAP.i1-Q4_K_S.gguf · Exact GGUF file →Historical first-MiB hash only; not a full-file identity.

Q5_K_M · 13.99 GB

mradermacher · recorded source

gemma-4-19b-a4b-it-REAP.i1-Q5_K_M.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.