Step 3.5 Flash
StepFun's 196B Apache 2.0 agent model from February 2026: 11B active, 256K context. Legacy, but still widely downloaded.
In short. Step 3.5 Flash is a text-only open model from the Chinese company StepFun, with about 200 billion parameters of which 11 billion work on each token. It was built for reasoning, coding and agents and handles up to 256K tokens. StepFun has since replaced it with Step 3.7 Flash, which adds image input, but this version is still downloaded a lot.
Summary
StepFun published Step 3.5 Flash on Hugging Face on 1 February 2026 under Apache 2.0, with a technical report on arXiv. It is a sparse mixture of experts with 196B parameters (199.4B on Hugging Face) that activates 11B per token. It uses a 3-to-1 mix of sliding-window and full attention layers to reach a 256K context cheaply, and 3-way multi-token prediction for speed, which StepFun says gives 100 to 300 tokens per second in typical use and up to 350 on single-stream coding. The model reads and writes text only. StepFun later released Step 3.7 Flash, with the same size plus image input, so we list this one as a legacy model.
In StepFun's own table, which shows reproduced scores for some rivals, Step 3.5 Flash scores 74.4 on SWE-bench Verified (DeepSeek V3.2 73.1, GLM-4.7 73.8, MiniMax M2.1 74.0), 51.0 on Terminal-Bench 2.0 (DeepSeek V3.2 46.4, GLM-4.7 41.0), 88.2 on tau2-Bench, 86.4 on LiveCodeBench v6, 97.3 on AIME 2025 and 98.4 on HMMT 2025 (Feb.), and 51.6 on BrowseComp, or 69.0 with its context manager. StepFun also estimates a decoding cost at 128K context that is far lower than those of DeepSeek V3.2, Kimi K2 Thinking and GLM-4.7. All scores are StepFun's own, and the models it compares against are from late 2025 and early 2026.
The BF16 weights are 400.4 GB and the FP8 weights 207.2 GB. StepFun's GGUF file at Q4_K_S is 111.5 GB, and the card lists 120 GB of memory as the minimum and 128 GB as recommended, on a Mac Studio, an NVIDIA DGX Spark or an AMD Ryzen AI Max+ 395 machine, with mainline llama.cpp. StepFun also published base and mid-training checkpoints. In the 30 days to 11 November 2026, Hugging Face counted about 179,600 downloads of the main repository, which makes it the most downloaded StepFun language model despite being superseded.
Good for
- Local or on-premise text agents on a 128 GB Mac or mini PC, using the 111.5 GB Q4_K_S file
- Reasoning, coding and search agents that need a long 256K context but not images
- Researchers who want the published base and mid-training checkpoints
- Existing deployments that already run it and are not ready to move to Step 3.7 Flash
Watch out for
- It is a legacy model, and Step 3.7 Flash has the same size with image input and, by StepFun's account, better agent scores
- All scores are StepFun's own, and some of the rival scores are StepFun's own reproductions
- Local use needs about 120 GB of memory
- It reads text only
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the GGUF Q4_K_S variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Step-3.5-Flash (BF16) BF16 | 199B 11B active | 459 GB | 229 GB | 138 GB | Big-memory desktop |
| Step-3.5-Flash-FP8 FP8 | 199B 11B active | 459 GB | 229 GB | 138 GB | Big-memory desktop |
| GGUF Q4_K_S INT4 | 199B 11B active | 459 GB | 229 GB | 120 GB | Big-memory desktop |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 120 GB of GPU memory.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
Frequently asked questions
- What is Step 3.5 Flash?
- Step 3.5 Flash is a text-only open model from the Chinese company StepFun, with about 200 billion parameters of which 11 billion work on each token. It was built for reasoning, coding and agents and handles up to 256K tokens. StepFun has since replaced it with Step 3.7 Flash, which adds image input, but this version is still downloaded a lot.
- Can I use Step 3.5 Flash commercially?
- Yes. Step 3.5 Flash is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Step 3.5 Flash open source?
- Step 3.5 Flash is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Step 3.5 Flash on my own computer?
- Only on a machine with a lot of memory. The smallest variant, GGUF Q4_K_S, needs about 120 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
- When was Step 3.5 Flash released?
- Step 3.5 Flash was released by StepFun on February 1, 2026.
- How long a context window does Step 3.5 Flash have?
- Step 3.5 Flash supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Step 3.5 Flash: Fast Enough to Think. Reliable Enough to Act. (StepFun) static.stepfun.com
- Step-3.5-Flash model card (Hugging Face) huggingface.co
- Step 3.5 Flash technical report (arXiv 2602.10604) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Step 3.7 Flash StepFun | 201B 11B active | 256K | Apache 2.0 (Permissive) | 120 GB Big-memory desktop | |
| Step 5 Preview StepFun | 604B 27B active | 1M | StepFun 1.0 (Conditional) | 417 GB Big-memory desktop | |
| DeepSeek-V3.2 DeepSeek | 685B 37B active | 160K | MIT (Permissive) | 473 GB Big-memory desktop | |
| GLM-5.2 Z.ai (Zhipu AI) | 753B 40B active | 1M | MIT (Permissive) | 520 GB GPU server |