Gemma 3 4B
Google · open weights · writing benchmark
The short version
Gemma 3 4B lands in the bottom third of ToneBench at #166 of 167, with a writing Elo of -187 and an overall score of 43.5 out of 100. Those numbers come from it writing its own version of the scripts behind 10 of my What's AI videos, five times each, with every draft scored blind against mine (how the scoring works).
For scale, my own scripts score 90.9 through the same judges and rubric. That's the real bar here, not a literal 100. It sits about 47 points under it.
Its Elo comes from comparing its average on each script with every other model's average on the same script, with gaps inside the run-to-run noise counted as draws. The 95% range is -220 to -155. That range overlaps 2 other configurations, ranked #164 to #165, so its exact spot inside that band is too close to call on this board. What the ranges can and can't tell you.
It's under the board average on every metric. The least-bad is Length, still about 24 points under. It's furthest behind on Cues, about 42 points under the board average.
The format changes the result. Its best score was on the founder announcement / personal origin story (52.3) and its weakest on the engineering process walkthrough / presentation adaptation (34.5), about 18 points apart. If most of your scripts look like one of those, weigh that score over the average.
It's also swingy: its overall moves ± 6.3 between runs of the same script, against ± 2.7 for the typical model, so the same prompt can give you anything from rough to unusable.
It costs about $0.001 per script. Cohere Command A+ (05-2026) scores far higher and is free on the route we used, so on value this isn't the pick. For full scripts, I wouldn't use it.
If you're on Google anyway, Gemini 3.1 Pro (default) is worth paying about 14 cents more per script.
Among open-weights models, it's 52nd of 53. The best open-weights config is GLM-5.3 at #15, and everything above it is closed, so open weights still trail the frontier on this kind of writing.
Skill profile
Read the shape: the further a corner reaches, the stronger the model is on that metric. The dashed line is the board average and the dotted line is the board's best score on each metric, so a corner outside the dashed line beats the average there. Hover a point for the exact numbers.
Metric by metric
Nine writing metrics, each scored 0–100 and blended into the overall by the editorial weight shown next to it. Rank is out of all 167 current-ranked models. The thick bar is this model; the thin line above it is the field's best on that metric (darker) and the one below is the field's average (lighter), on the same scale. The small ± is how much the score moves from run to run, and the ? next to each name says what the metric rewards.
Script by script
The same 10 real scripts every current-ranked model writes, each scored on its own as the mean of 5 runs. Different formats stress different skills: a model can nail a tight explainer and still stumble on a personal story, so look for the format closest to what you write.
Consistency: how much it moves between runs
A model whose drafts swing from run to run is harder to use than its average suggests. Every score on this page is the mean of 5 runs per script, and the ± is the run-to-run standard deviation.
Its most volatile metric is Hook: ± 6.6 between runs, when the typical model moves ± 2.0, so two drafts of the same script can score quite differently there.
| Metric | This model | Board median | Verdict |
|---|---|---|---|
| Tone ? | 44.5 ± 3.5 | ± 1.7 | swingier than most |
| Craft ? | 48.5 ± 3.5 | ± 1.5 | swingier than most |
| Substance ? | 39.8 ± 4.3 | ± 2.0 | swingier than most |
| Flow ? | 43.0 ± 3.1 | ± 1.8 | swingier than most |
| YouTube ? | 36.1 ± 3.2 | ± 2.9 | typical |
| Hook ? | 42.7 ± 6.6 | ± 2.0 | swingier than most |
| Length ? | 46.5 ± 8.3 | ± 8.6 | typical |
| Slop ? | 59.4 ± 4.5 | ± 2.1 | swingier than most |
| Cues ? | 35.3 ± 6.3 | ± 3.2 | swingier than most |
Cost, time and tokens
Logged on every run. None of it touches the writing score, but it's often what decides whether a model fits your pipeline at all. Time is the average time to get one accepted script, retries included, and cost is priced per script at list price.
What each judge scored it
One judge can have taste of its own, which is why there are three, from three model families, each scoring the same 50 stored drafts blind with the identical rubric. The overall of 43.5 combines their scores with the parts computed in code (length, and half of the slop score). The highest and lowest judge are 11.9 points apart on its overall. That's wider than typical (board median 6.3): GPT-5.6 Sol (medium) rates it noticeably higher than Claude Opus 5, so the consensus sits between two real opinions. How the panel works.
How we ran Gemma 3 4B
- Accessed viaOpenRouter · model page ↗
- Exact model / route id
google/gemma-3-4b-it(provider default reasoning) - Generations5 independent runs × 10 scripts = 50 drafts, stored immutably after generation
- Run dates2026-09-03
- Judges3-judge panel: Claude Opus 5, high effort (Anthropic) + DeepSeek V4.1 Flash (OpenRouter) + GPT-5.6 Sol, medium reasoning (Codex CLI). All three score every draft independently with the identical rubric, blind to which model wrote it, and each one has to quote evidence before it scores a metric. Seven metrics are the plain average of the three judges; Length is computed from the spoken word count, and Slop is half computed (the EQ-Bench slop-score pipeline plus our banned-phrase scan) and half judged. The Elo is cross-checked with a Bradley-Terry refit (methodology)
- Sampling and promptprovider-default sampling, no fine-tuning, no system-prompt changes. Every current-ranked model gets the same prompt, about 11,000 tokens (the count shifts a bit with each model's tokenizer): a style guide built from my voice and storytelling guides, the target length, and the brief with its research packet. What changes from row to row is the model configuration
- What stays privatethe exact rubric and any non-public reference material. Scripts are normally benchmarked before their video goes out, which lowers the odds a model has seen mine, though it doesn't make that impossible (why)
Frequently asked questions
How good is Gemma 3 4B at writing?
It's #166 of 167 on ToneBench, with a writing Elo of -187 and an overall score of 43.5/100. It wrote its own version of the scripts behind 10 of my What's AI videos 5 times each, and three judges from three model families scored every draft blind against mine. Length and half of the slop score are computed in code.
Is Gemma 3 4B the best Google model for writing?
No. It's last of 17 Google configs; Gemini 3.1 Pro (default) is about 1652 Elo ahead.
Is Gemma 3 4B good value for the money?
It costs about $0.001 per script. Cohere Command A+ (05-2026) scores higher and costs less, so it isn't the value pick.
What are Gemma 3 4B's strengths and weaknesses?
It's under the board average on every metric; the least-bad is Length (46.5, -24.3) and it's furthest behind on Cues (35.3, -41.6). The nine-metric breakdown on this page shows the board average and board best for each.
How was Gemma 3 4B evaluated?
Through OpenRouter, using the exact model/route id google/gemma-3-4b-it, first run on 2026-09-03. Each script ran 5 times with provider-default sampling and no fine-tuning. The 'How we ran it' section and the methodology page have the rest.
Is Gemma 3 4B open source?
Yes, it's an open-weights model. Among open-weights models it's 52nd of 53 for writing.