1 tested weight format · Explore results, then inspect the answers.
Q4_0 · 4.59 GB download
Links follow the selected quant and configuration.Q4_0 · 4.59 GB of weights · rocm · context 32768 · GPU layers 99
Medium · 5–under 20B · dense · 7.5B total
Memory fit has not been verified for this configuration.
completed
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.
30.2 / 100
Nine categories, equally weighted · 9/9 eligible
Complete coverage for this setup. Quality only; speed is separate. How the index works →
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.
Each suite has its own metric and fixed task bank. The balanced index weights categories equally; individual scores and evidence remain available below.
| Test suite | Thinking off | Details |
|---|---|---|
| Everyday coding tasksTask pass rate |
| |
| Larger practical tasksMean task credit |
| |
| Logic and number puzzlesMean puzzle credit |
| |
| Published hard problemsMean answer credit |
| |
| Following exact instructions · original campaignStrict prompt compliance |
| |
| Reading without guessingExact-match accuracy |
| |
| Calculating from a passageNumeric exact-match accuracy |
| |
| Reading research papersAnswer token F1 |
| |
| Resisting incorrect correctionsAll-branches pass rate |
| |
| Following exact instructions · 128 GB campaignStrict prompt compliance | Not planned |
This analysis covers this configuration across its recorded thinking conditions and test suites.
Use Gemma 4 E4B with thinking on for small project edits and logic puzzles. It also answered passage questions with short task times in these runs. Research-paper answering remains weak. Larger practical work had mixed outcomes across the fully scored eligible task bank, so small-task successes should not be taken as a promise of dependable unattended project work.
Logic and number puzzles
Thinking on. Medium · 5–under 20B · dense. Rank 1 of 9 in this recorded comparison.
| Tasks passed | 16 of 16 (100.0%) |
|---|---|
| Test coverage | 16 of 16 tasks scored |
| Average score | 82.4/100 · Mean puzzle credit |
|---|---|
| Test coverage | 34 of 34 tasks scored |
| Task outcomes | 28 full credit · 0 partial · 6 no credit |
| Tasks passed | 14 of 16 (87.5%) |
|---|---|
| Test coverage | 16 of 16 tasks scored |
| Tasks timed | 16 tasks |
| Typical time per task (median) | 4.2 seconds |
| Average score | 34.7/100 · Answer token F1 |
|---|---|
| Test coverage | 8 of 8 tasks scored |
| Task outcomes | 0 full credit · 6 partial · 2 no credit |
| Average score | 54.5/100 · Mean task credit |
|---|---|
| Test coverage | 11 of 11 tasks scored |
| Task outcomes | 6 full credit · 0 partial · 5 no credit |
Original analysis generated by GPT Astra · · Edited by Codex; agent reviewed by Codex · publication authorized by user
Codex reviewed this assessment against a newer evidence edition. The original analyst, generation date and source draft remain recorded separately from the editorial revision.
Generation and editorial authorship remain attached to this revision. Changing the default generator does not change this attribution.
Handles small project edits and tight wording, length and formatting requirements well with thinking on. It passed every small-coding task, though that does not establish larger-project ability.
| Tasks passed | 16 of 16 (100.0%) |
|---|---|
| Test coverage | 16 of 16 tasks scored |
| Tasks passed | 21 of 23 (91.3%) |
|---|---|
| Test coverage | 23 of 23 tasks scored |
Keeps source-backed answers when incorrectly challenged and updates them when the source changes. With thinking on, it passed every tested correction scenario.
| Tasks passed | 12 of 12 (100.0%) |
|---|---|
| Test coverage | 12 of 12 tasks scored |
Research-paper answering is weak even with thinking: answer matching scores were low, and no answer matched fully. Treat its paper answers as starting points requiring verification, not dependable research conclusions.
| Average score | 34.7/100 · Answer token F1 |
|---|---|
| Test coverage | 8 of 8 tasks scored |
| Task outcomes | 0 full credit · 6 partial · 2 no credit |
Difficult math and reasoning remain uneven with thinking on. Larger app-building and investigative tasks also produced mixed outcomes across the fully scored eligible task bank; supervised use remains appropriate.
| Average score | 46.7/100 · Mean answer credit |
|---|---|
| Test coverage | 11 of 11 tasks scored |
| Task outcomes | 5 full credit · 2 partial · 4 no credit |
| Average score | 54.5/100 · Mean task credit |
|---|---|
| Test coverage | 11 of 11 tasks scored |
| Task outcomes | 6 full credit · 0 partial · 5 no credit |
Keep thinking on for everyday use: it improves small-project coding at a modest extra time cost, while passage answering and instruction following improve and take less average time. Turning it off is not a dependable speed shortcut.
| Thinking off → on | 68.8% · Task pass rate → 100.0% · Task pass rate |
|---|---|
| Paired test coverage | 16 of 16 tasks compared in both modes |
| Tasks timed in both modes | 16 pairs |
| Score change with thinking on | +31.3 score points |
| Average time change with thinking on | 4.9 seconds slower |
| Thinking off → on | 25.0% · Exact-match accuracy → 87.5% · Exact-match accuracy |
|---|---|
| Paired test coverage | 16 of 16 tasks compared in both modes |
| Tasks timed in both modes | 16 pairs |
| Score change with thinking on | +62.5 score points |
| Average time change with thinking on | 40.8 seconds faster |
| Thinking off → on | 17.4% · Strict prompt compliance → 91.3% · Strict prompt compliance |
|---|---|
| Paired test coverage | 23 of 23 tasks compared in both modes |
| Tasks timed in both modes | 23 pairs |
| Score change with thinking on | +73.9 score points |
| Average time change with thinking on | 33.4 seconds faster |
With thinking on, it narrowly leads other tested medium-sized dense models on puzzle scores and ties the best small-coding result. Difficult math and reasoning are less competitive, sitting between substantially stronger and weaker alternatives.
Logic and number puzzles
Thinking on. Medium · 5–under 20B · dense. Rank 1 of 9 in this recorded comparison.
Everyday coding tasks
Thinking on. Medium · 5–under 20B · dense. Tied at rank 1 of 9 in this recorded comparison.
Published hard problems
Thinking on. Medium · 5–under 20B · dense. Rank 4 of 9 in this recorded comparison.
| Average score | 82.4/100 · Mean puzzle credit |
|---|---|
| Test coverage | 34 of 34 tasks scored |
| Task outcomes | 28 full credit · 0 partial · 6 no credit |
| Tasks passed | 16 of 16 (100.0%) |
|---|---|
| Test coverage | 16 of 16 tasks scored |
| Average score | 46.7/100 · Mean answer credit |
|---|---|
| Test coverage | 11 of 11 tasks scored |
| Task outcomes | 5 full credit · 2 partial · 4 no credit |
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.
Other mode, this model only: thinking off · 14.9/100 · Mean puzzle credit.
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.
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.
Several models reached full marks; this test cannot separate their top scores.
Other mode, this model only: thinking off · 68.8% · Task pass rate.
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.
Answer selected difficult math and reasoning questions from LiveBench.
Mean answer 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.
Other mode, this model only: thinking off · 9.4/100 · Mean answer credit.
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.
With thinking on, reading accuracy tied Qwen3.5 9B and was one answer behind Qwen3.5 4B, while median task time was much shorter than both.
Speed: Reading without guessing
Thinking on. Matching test-machine label and execution settings; spans model sizes. Timings include unsuccessful tasks, so read them alongside quality.
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.
| Model | Typical task time | Generation speed | Tasks passed |
|---|---|---|---|
| Gemma 4 E4BThis model | 4.2 seconds16 of 16 tasks timed | 73.0 tokens/s | 14 of 1616 of 16 scored |
| Qwen3.5 4B | 27.2 seconds16 of 16 tasks timed | 76.9 tokens/s | 15 of 1616 of 16 scored |
| Qwen3.5 9B | 37.8 seconds16 of 16 tasks timed | 55.6 tokens/s | 14 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
Confirmed context-budget failures receive zero task-completion credit and count as failures to answer, not incorrect submitted answers. Answer-only averages exclude these no-answer outcomes; completed same-budget retries are scored normally. This follow-up is selected after observing failures.
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.
Publication redactions restrict reader inspection of some examples and paper material. They do not establish that the model lacked those inputs.
Source limitation
Some selected evidence has recorded publication redactions.
One weight format is recorded for this model. Additional tested formats will appear here when available.
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.
16/16 matched scored questions · 16/16 matched timed questions
Thinking on took 4.9 seconds more per matched timed question, on average.
| Matched measurement | Recorded result |
|---|---|
| Average credit, off → on | 68.8% → 100.0% |
| Mean paired credit change | +31.3 percentage points |
| Question outcomes | 5 improved · 11 unchanged · 0 worse |
| Mean added task time | +4.9s |
| Mean task time, off → on | 7.5s → 12.4s |
| Median task time, off → on | 5.4s → 11.3s |
| Median of paired time changes | +4.1s |
| Total added task time | +78.9s |
| Ratio of summed times, on / off | 1.7× |
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 →Repeated runs are a separate experiment and never add votes to the baseline score.
48/48 repeats scored · 35 full-credit repeats.
16/16 questions have every planned repeat scored. Of those, 8 earned full credit on every repeat.
Find information in a small project, edit files and fix bugs.
Each segment counts tasks, not credit points. Partial credit varies by task.
Scoring and fixed question bank →
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
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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThe 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 · primaryThese are starting points for inspection, not claims that one setup is better. Runtime and placement differences remain visible.
Verified requested on/off conditions of the same configuration, matched on the recorded test protocol.
Requested thinking on · Complete compared coverageCompare these configurations →
Another checkpoint in the same total-parameter band and architecture category, with matching task controls.
Requested thinking off · Complete compared coverageCompare these configurations →
ggml-org · recorded source
gemma-4-E4B-it-Q4_0.gguf · Exact GGUF file →Historical first-MiB hash only; not a full-file identity.
Choose a suite to inspect its exact available questions, references and model answers.