Intel / post

STEP3-VL-10B claims 200B-class multimodal results at 10B parameters

Source: x.com
Date: Published Jan 20, 2026

Context

StepFun released STEP3-VL-10B, a 10-billion-parameter open vision-language model, reporting scores that beat GLM-4.6V, a 106-billion-parameter model, and Qwen3-VL, a 235-billion-parameter model, on reasoning benchmarks such as MMMU and MathVision. StepFun attributes this to 1.2 trillion tokens of full-parameter , over 1,400 reinforcement-learning iterations refining its reasoning, and a technique it calls PaCoRe that allocates compute dynamically during rather than using a fixed amount for every query.

These benchmark comparisons are StepFun's own reported results rather than an independent leaderboard placement, and the announcement does not specify how GLM-4.6V or Qwen3-VL were run for the comparison. StepFun published a paper, weights for both a base and a thinking variant, and hosted playgrounds on Hugging Face and ModelScope, letting outside groups reproduce the comparison on their own hardware.

Terms in this piece · Glossary

multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.

Read the source x.com