Step3 VL 10B

StepFun · Released Jan 20, 2026
DFO score #365 / 698
12.7 leaderboard score
Speed
– Not reported
Input Price
– Not priced
Output Price
– Not priced
Context
– Not reported

Analysis Summary

StepFun's Step3 VL 10B is a compact vision-language model by its name, with supplied benchmark results showing modest general reasoning and useful performance on selected knowledge and instruction-following tasks. Its tau2 result is low, and no coding index, long-context score, pricing, or speed information is provided. The benchmark record does not establish a strong agentic profile.

It may warrant a limited evaluation for workloads that need a smaller model and align with the measured knowledge-task performance. The data is not sufficient to recommend it for complex client-facing analysis, autonomous tool use, or production coding. Since cost and operational characteristics are unknown, compare it with alternatives on the actual tasks, including any visual inputs the team intends to process. Until that evaluation demonstrates a clear advantage, treat it as an experimental specialist rather than a default workflow model.

Assessed September 23, 2026

Editorial notes

Step3 VL 10B has some useful knowledge and instruction-following results, but low agentic performance and limited evidence on coding, context, and pricing restrict its production fit.

The DFO score ranks models on the Artificial Analysis Intelligence Index and is refreshed as new models launch. The write-up and verdict are our own. Feedback?

DFO Verdict

Step3 VL 10B has some useful knowledge and instruction-following results, but low agentic performance and limited evidence on coding, context, and pricing restrict its production fit.

#365 of 698 overall Down 2 this week

Benchmarks

Intelligence Index 7.7 #284 of 436 · Artificial Analysis
Coding Index Not published yet Artificial Analysis
Agentic Index Not published yet Artificial Analysis
Creative Writing Not rated Arena rating
Instruction Following Not rated Arena rating
GPQA Diamond 69%
HLE 10.2%
SciCode 31.1%
TerminalBench Hard 5.3%
τ²-Bench 16.1%
IFBench 50.2%

Artificial Analysis data refreshed Oct 12, 2026. Arena ratings from the Oct 2, 2026 text leaderboard, with style control. Bars: magenta = reasoning, ink = coding and agents, cyan = instructions and long context.

Artificial Analysis has not published coding or agentic results for Step3 VL 10B yet. Until it does, the model sits below tested models on the Coding and AI Agents lists.

How Step3 VL 10B compares

Step3 VL 10B ranks #284 of 436 AI models we track for overall intelligence. Step3 VL 10B is currently free to use via OpenRouter.

Position in the field
DFO score: ahead of 48% of models #361
worst in fieldmedianbest in field
Price vs frontier peers · $ per 1M tokens
Step3 VL 10B $0.00 in $0.00 out
Anthropic: Claude Opus 5.5 $4.00 in $20.00 out
Anthropic: Claude Sonnet 5.5 $2.00 in $10.00 out
Anthropic: Claude Fable 5.1 $10.00 in $50.00 out

Dark bar = input · light bar = output, scaled to the priciest peer.

Context window vs peers · tokens

1M tokens ≈ 8 full-length novels or ~2,500 pages of business documents in a single request.

Frequently asked questions about Step3 VL 10B

How much does Step3 VL 10B cost?

Step3 VL 10B is currently available for free via OpenRouter.

Who created Step3 VL 10B?

Step3 VL 10B is developed by StepFun and was released on January 20, 2026.

© 2026 Design for Online Ltd. Registered in England and Wales No. 10328553. VAT Registered. Design for Online® and Forerunner® are registered trademarks.