AgentBench / results
← All models

Phi-4 Mini Reasoning 3.8B Q4_K_M

1 tested weight format · Explore results, then inspect the answers.

Choose a tested configuration

Get this tested model

Q4_K_M · 2.49 GB download

Links follow the selected quant and configuration.

16 held · thinking control unverified

Setup, memory & provenance · Q4_K_M

Q4_K_M · 2.49 GB of weights · vulkan · context 16384 · GPU layers 99

Small · under 5B · dense · 3.8B total

Memory fit has not been verified for this configuration.

partial Integration check held.

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

0–100 / 100

Nine categories, equally weighted · 0/9 eligible

Possible score range with unscored tasks receiving zero to full credit. Incomplete results remain unranked; this is not a confidence interval. 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 rateNot scored0/16 scored · PreviewNo measured time
Larger practical tasksMean task creditNot scored0/11 scored · PreviewNo measured time
Logic and number puzzlesMean puzzle credit
Average credit
—
Full-credit rate
0/0 · —
Answers completed
0/0 · —
Full credit / completed
0/0 · —
34 technical / untested / held34 unscored; excluded from these rates.
0/34 scored · PreviewNo measured time
Published hard problemsMean answer credit
Average credit
—
Full-credit rate
0/0 · —
Answers completed
0/0 · —
Full credit / completed
0/0 · —
11 technical / untested / held11 unscored; excluded from these rates.
0/11 scored · PreviewNo measured time
Following exact instructions · original campaignStrict prompt compliance
Pass rate
0/0 · —
Answers completed
0/0 · —
Correct / completed
0/0 · —
23 technical / untested / held23 unscored; excluded from these rates.
0/23 scored · PreviewNo measured time
Reading without guessingExact-match accuracy
Pass rate
0/0 · —
Answers completed
0/0 · —
Correct / completed
0/0 · —
16 technical / untested / held16 unscored; excluded from these rates.
0/16 scored · PreviewNo measured time
Calculating from a passageNumeric exact-match accuracy
Pass rate
0/0 · —
Answers completed
0/0 · —
Correct / completed
0/0 · —
7 technical / untested / held7 unscored; excluded from these rates.
0/7 scored · PreviewNo measured time
Reading research papersAnswer token F1
Average credit
—
Full-credit rate
0/0 · —
Answers completed
0/0 · —
Full credit / completed
0/0 · —
8 technical / untested / held8 unscored; excluded from these rates.
0/8 scored · PreviewNo measured time
Resisting incorrect correctionsAll-branches pass rate
Pass rate
0/0 · —
Answers completed
0/0 · —
Correct / completed
0/0 · —
12 technical / untested / held12 unscored; excluded from these rates.
0/12 scored · PreviewNo measured time
Following exact instructions · 128 GB campaignStrict prompt complianceNot planned

Model analysis

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

Reserve this configuration for low-stakes experimentation, not dependable reading, precise formatting or reasoning. With thinking requested off, completed text tests were weak, especially passage answers, calculations, research-paper answers and handling source-backed corrections. Coding and longer tool-based work remain unevaluated previews, so a practical coding niche is unproven. Whether enabling thinking helps is also unknown.

Evidence for this judgment
Following exact instructions · Thinking off
Tasks passed8 of 23 (34.8%)
Test coverage23 of 23 tasks scored
Tasks timed23 tasks
Typical time per task (median)15.2 seconds
Explore these test results →

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

Logic and number puzzles · Thinking off
Average score26.5/100 · Mean puzzle credit
Test coverage34 of 34 tasks scored
Task outcomes9 full credit · 4 partial · 21 no credit
Explore these test results →

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

Published hard problems · Thinking off
Average score10.5/100 · Mean answer credit
Test coverage11 of 11 tasks scored
Task outcomes1 full credit · 2 partial · 8 no credit
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 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 →

Resisting incorrect corrections · Thinking off
Tasks passed0 of 12 (0.0%)
Test coverage12 of 12 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 →

Everyday coding tasks · Thinking off
Tasks passed0 of 0 (score unavailable)
Test coverage0 of 16 tasks scored
Explore these test results →

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

Larger practical tasks · Thinking off
Average scoreNot measured
Test coverage0 of 11 tasks scored
Task outcomes0 full credit · 0 partial · 0 no credit
Explore these test results →

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

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
Agent tasks remain held after completed diagnostic responses invented error text or produced malformed calls. Those diagnostics establish a local tool-protocol limitation, not a scored agent-task failure rate. This is a separate checkpoint from Phi Mini Instruct; no matched thinking-off counterpart is established.

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
2026-10-13T04:05:43.702056Z
Review time
2026-10-22T07:26:27.209037+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-73f4b2d4adb49cc7b187b6a5b9e67977c4ea2ffe2b52c30c4946f600560023fc
Analysis input hash
66963be65f6b3c54bf318805c546edd75d9abb25c6cf795d91dcd24a49ce8059
Original analysis input hash
36373681b74986c3160fdf07063bdf40e7f6df1091aaa2eed4a51bf4a053f845
Editorial evidence edition
dataset-5a6f96ec56c3d240cc42fa23147c1495da1be71f3ef0722c0e5f8c4f5e126a83
Original summary
summary-87fb3196237efc9a9673af228ff8f021fca9b4768588f7b811086807b7214ca1
Summary revision
summary-20369a16cc7a3f31c6c6abecf7c4af6bdbbb9b84d0b8e97c909ea4c9a8da97de

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

Contributing analyses (1)
  • GPT Astra ·
    Analysis stage
    direct
    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
    1e3b8e0d983e10ab476b40fea9afe8ed53d31eba5297236ceb6412be845682f7
    Summary revision
    summary-549b75ee552452efb4b256b4a2898772635b40a6ef4359b3e81680b660ec9b89
    Input summary revisions
    Direct evidence analysis

Struggles with

Source-based answers are a major weakness: passage extraction and calculations failed throughout, paper-answer matching was extremely poor, and the model could not satisfy the combined requirements for resisting incorrect challenges and accepting changed evidence.

Evidence for this judgment
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 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 →

Resisting incorrect corrections · Thinking off
Tasks passed0 of 12 (0.0%)
Test coverage12 of 12 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 →

Exact wording and formatting need close checking. Strict instruction compliance was low, while puzzle solving and difficult math/reasoning were also weak across these tests.

Evidence for this judgment
Following exact instructions · Thinking off
Tasks passed8 of 23 (34.8%)
Test coverage23 of 23 tasks scored
Explore these test results →

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

Logic and number puzzles · Thinking off
Average score26.5/100 · Mean puzzle credit
Test coverage34 of 34 tasks scored
Task outcomes9 full credit · 4 partial · 21 no credit
Explore these test results →

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

Published hard problems · Thinking off
Average score10.5/100 · Mean answer credit
Test coverage11 of 11 tasks scored
Task outcomes1 full credit · 2 partial · 8 no credit
Explore these test results →

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

Compared with other models

With thinking requested off, strict-instruction success ties at the low end of the tested small dense group. Every alternative in the matched timing cohort finishes sooner, with equal or higher success. These results give it no useful speed-quality advantage for instruction-focused work.

Evidence for this judgment
Following exact instructions · Thinking off
Tasks passed8 of 23 (34.8%)
Test coverage23 of 23 tasks scored
Tasks timed23 tasks
Typical time per task (median)15.2 seconds
Explore these test results →

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

Following exact instructions
Thinking off. Small · under 5B · dense. Tied at rank 8 of 9 in this recorded comparison.

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

Following exact instructionsThinking off · Small · under 5B · dense

Keep to rules about wording, length, language and formatting.

Strict prompt compliance · 0–100% · higher is better
23 of 23 tasks scored for each model.

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

  1. #1 LFM2.5 1.2B Q4_01.2B total parameters87.0%
    12 tasks more passed
  2. #2 Qwen3.5 4B4.2B total parameters82.6%
    11 tasks more passed
  3. =3 Gemma 4 E2B Q4_04.6B total parameters69.6%
    8 tasks more passed
  4. =3 Granite 4.0 Micro 3.4B Q4_03.4B total parameters69.6%
    8 tasks more passed
  5. #5 Qwen3.5 2B Q4_K_M1.9B total parameters65.2%
    7 tasks more passed
  6. #6 Phi-4 Mini Instruct 3.8B Q4_K_M3.8B total parameters60.9%
    6 tasks more passed
  7. #7 SmolLM3 3B Q4_K_M3.1B total parameters52.2%
    4 tasks more passed
  8. =8 Phi-4 Mini Reasoning 3.8B Q4_K_M3.8B total parameters34.8%
    This model
  9. =8 Qwen3.5 0.8B Q4_00.8B total parameters34.8%
    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.

Comparison saved with the 2026-10-13 assessment. Task-specific results.

Speed and task success: Following exact instructionsThinking 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_01.7 seconds23 of 23 tasks timed319.0 tokens/s20 of 2323 of 23 scored
Qwen3.5 0.8B Q4_02.9 seconds23 of 23 tasks timed248.5 tokens/s8 of 2323 of 23 scored
Ling 3.0 Tiny Q4_03.3 seconds23 of 23 tasks timed138.9 tokens/s14 of 2323 of 23 scored
Mellum 2.1 12B-A2.5B Thinking3.6 seconds23 of 23 tasks timed140.2 tokens/s14 of 2323 of 23 scored
Granite 4.0 Micro 3.4B Q4_03.6 seconds23 of 23 tasks timed111.7 tokens/s16 of 2323 of 23 scored
SmolLM3 3B Q4_K_M3.7 seconds23 of 23 tasks timed129.6 tokens/s12 of 2323 of 23 scored
Phi-4 Mini Instruct 3.8B Q4_K_M3.8 seconds23 of 23 tasks timed108.7 tokens/s14 of 2323 of 23 scored
Nemotron Nano 9B v2 Q4_04.9 seconds23 of 23 tasks timed53.6 tokens/s19 of 2323 of 23 scored
Qwen2.5 Coder 7B Instruct Q4_K_M5.0 seconds23 of 23 tasks timed65.1 tokens/s13 of 2323 of 23 scored
Granite 4.2 8B Q4_06.9 seconds23 of 23 tasks timed57.7 tokens/s18 of 2323 of 23 scored
Ornith 1.5 9B Q4_K_M7.3 seconds23 of 23 tasks timed54.8 tokens/s12 of 2323 of 23 scored
Gemma 4 12B QAT Q4_09.4 seconds23 of 23 tasks timed37.8 tokens/s20 of 2323 of 23 scored
Ministral 3 8B Instruct Q4_K_M10.2 seconds23 of 23 tasks timed54.2 tokens/s14 of 2323 of 23 scored
Phi-4 Mini Reasoning 3.8B Q4_K_MThis model15.2 seconds23 of 23 tasks timed105.8 tokens/s8 of 2323 of 23 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 2026-10-13 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

Coding and repeat runs remain incomplete previews, not measured failures. Reasoning control is unverified, so these records do not establish whether enabling thinking improves quality or task time.

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.

Everyday coding tasks · Thinking off · sampled repetitions
Repetitions scored0 repetitions
Repetitions planned48 repetitions
Explore these test results →

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

Everyday coding tasks · Thinking on · sampled repetitions
Repetitions scored0 repetitions
Repetitions planned48 repetitions
Explore these test results →

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

Tool-based testing is held after diagnostics invented error messages or produced malformed calls. This establishes a local tool-protocol problem, not a task-failure rate. The separate Mini Instruct checkpoint is not a matched thinking-off diagnostic counterpart.

Evidence for this judgment

Source limitation
Agent tasks remain held after completed diagnostic responses invented error text or produced malformed calls. Those diagnostics establish a local tool-protocol limitation, not a scored agent-task failure rate. This is a separate checkpoint from Phi Mini Instruct; no matched thinking-off counterpart is established.

The publication exposes only a source link for the paper example's full paper-containing input, limiting reader inspection. This does not establish that the model lacked the paper.

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

Source limitation
Qualitative examples are a deterministic sample; they do not establish whole-bank behavior.

Example 1: Reading research papers
Thinking off. Final result: no 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 →

How to interpret these assessments and their limitations →

One weight format is recorded for this model. Additional tested formats will appear here when available.

What changed in these results

Agent tasks remain held after completed diagnostic responses invented error text or produced malformed calls. Those diagnostics establish a local tool-protocol limitation, not a scored agent-task failure rate. This is a separate checkpoint from Phi Mini Instruct; no matched thinking-off counterpart is established.

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.

Sampled repetitions

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

Thinking on · temperature 0.6

The required reasoning control was not verified. Repetition quality is not assessed.

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

Everyday coding tasks: recorded outcomes

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

0 full credit · 0 partial credit · 0 zero credit

16 pending

Ranking unavailable: incomplete coverage, unverified reasoning control.

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.

Qwen3.5 4B

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

Requested thinking on · Exploratory: incomplete or uncertain evidence

Compare these configurations →

Selection rule and evidence IDsPolicy comparisons-v3 · Template same-parameter-architecture-matching-controls
Facts: fact-9d133fa304209ce92e03d207, fact-8471cd6afce688a52c7066ee
Matching candidate set: fact-8471cd6afce688a52c7066ee, fact-ed9e4adee7986e86f4bfc2e4, fact-4e0224af8f948a39b33d46fa, fact-bc54601b0f1841d544c224c5, fact-13bffa6d739b0531f72616b4, fact-7578ebf4ba24c76f5ad6ef40, fact-3a9b3fb80c44f1f53c33c065, fact-093ef2fab39b29e20dc389a1, fact-ee68991b61fb7db54bc8ee72

Tested weights

Q4_K_M · 2.49 GB

lmstudio-community · recorded source

Phi-4-mini-reasoning-Q4_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.