Model reference · open weights
Step-3.5-Flash-Midtrain is an open-weight language model from stepfun-ai. AxForge deploys and operates it for you on dedicated EU-owned hardware — with the licence handled where one is required.
Available as managed deployment — configured and operated for you on dedicated EU hardware, quoted per deployment.
What it is
| Maker | stepfun-ai |
|---|---|
| Type | Language models |
| Task | Text gen |
| Parameters (lead) | 197.8B |
| Context | 128k tokens |
| Runs with | transformers |
| Released | 2026-03-02 |
| Popularity | 152 downloads / month |
| Licence | Open weights |
About
Step 3.5 Flash (visit website) is our most capable open-source foundation model, engineered to deliver frontier reasoning and agentic capabilities with exceptional efficiency. We also open-sourced the training codebase (SteptronOss), with support for continue pretrain, SFT, RL (WIP), and evaluation (WIP), and will open-source the SFT data. Built on a sparse Mixture of Experts (MoE) architecture, it selectively activates only 11B of its 196B parameters per token. This "intelligence density" allows it to rival the reasoning depth of top-tier proprietary models, while maintaining the agility required for real-time interaction.
Deep Reasoning at Speed: While chatbots are built for reading, agents must reason fast. Powered by 3-way Multi-Token Prediction (MTP-3), Step 3.5 Flash achieves a generation throughput of 100–300 tok/s in typical usage (peaking at 350 tok/s for single-stream coding tasks). This allows for complex, multi-step reasoning chains with immediate responsiveness.
A Robust Engine for Coding & Agents: Step 3.5 Flash is purpose-built for agentic tasks, integrating a scalable RL framework that drives consistent self-improvement. It achieves 74.4% on SWE-bench Verified and 51.0% on Terminal-Bench 2.0, proving its ability to handle sophisticated, long-horizon tasks with unwavering stability.
Efficient Long Context: The model supports a cost-efficient 256K context window by employing a 3:1 Sliding Window Attention (SWA) ratio—integrating three SWA layers for every full-attention layer. This hybrid approach ensures consistent performance across massive datasets or long codebases while significantly reducing the computational overhead typical of standard long-context models.
Accessible Local Deployment: Optimized for accessibility, Step 3.5 Flash brings elite-level intelligence to local environments. It runs securely on high-end consumer hardware (e.g., Mac Studio M4 Max, NVIDIA DGX Spark), ensuring data privacy without sacrificing performance.
Step 3.5 Flash delivers performance parity with leading closed-source systems while remaining open and efficient.
Performance of Step 3.5 Flash measured across Reasoning, Coding, and Agentic Abilities. Open-source models (left) are sorted by their total parameter count, while top-tier proprietary models are shown on the right. xbench-DeepSearch scores are sourced from official publications for consistency. The shadowed bars represent the enhanced performance of Step 3.5 Flash using Parallel Thinking.
| Benchmark | # Shots | Step3.5 Flash (Base Midtrain) | Step3.5 Flash (Base) | MiMo‑V2 Flash (Base) | GLM‑4.5 (Base) | DeepSeek V3.1 (Base) | DeepSeekV3.2 (Exp Base) | Kimi‑K2 (Base) |
|---|---|---|---|---|---|---|---|---|
| # Activated Params | - | 11B | 11B | 15B | 32B | 37B | 37B | 32B |
| # Total Params | - | 196B | 196B | 309B | 355B | 671B | 671B | 1043B |
| General | ||||||||
| BBH | 3-shot | 87.3 | 88.2 | 88.5 | 86.2 | 88.2† | 88.7† | 88.7 |
| MMLU | 5-shot | 83.4 | 85.8 | 86.7 | 86.1 | 87.4† | 87.8† | 87.8 |
| MMLU‑Redux | 5-shot | 87.2 | 89.2 | 90.6 | - | 90.0† | 90.4† | 90.2 |
| MMLU‑Pro | 5-shot | 63.6 | 62.3 | 73.2 | - | 58.8† | 62.1† | 69.2 |
| HellaSwag | 10-shot | 91.0 | 90.2 | 88.5 | 87.1 | 89.2† | 89.4† | 94.6 |
| WinoGrande | 5-shot | 75.8 | 79.1 | 83.8 | - | 85.9† | 85.6† | 85.3 |
| GPQA | 5-shot | 43.8 | 41.7 | 43.5* | 33.5* | 43.1* | 37.3* | 43.1* |
| SuperGPQA | 5-shot | 41.2 | 41.0 | 41.1 | - | 42.3† | 43.6† | 44.7 |
| SimpleQA | 5-shot | 28.4 | 31.6 | 20.6 | 30.0 | 26.3† | 27.0† | 35.3 |
| Mathematics | ||||||||
| GSM8K | 8-shot | 88.9 | 88.2 | 92.3 | 87.6 | 91.4† | 91.1† | 92.1 |
| MATH | 4-shot | 65.7 | 66.8 | 71.0 | 62.6 | 62.6† | 62.5† | 70.2 |
| Code | ||||||||
| HumanEval | 3-shot | 67.0 | 81.1 | 77.4* | 79.8* | 72.5* | 67.7* | 84.8* |
| MBPP | 3-shot | 79.0 | 79.4 | 81.0* | 81.6* | 74.6* | 75.6* | 89.0* |
| HumanEval+ | 0-shot | 75.0 | 72.0 | 70.7 | - | 64.6† | 67.7† | - |
| MBPP+ | 0-shot | 62.4 | 70.6 | 71.4 | - | 72.2† | 69.8† | - |
| MultiPL‑E HumanEval | 0-shot | 63.0 | 67.7 | 59.5 | - | 45.9† | 45.7† | 60.5 |
| MultiPL‑E MBPP | 0-shot | 47.9 | 58.0 | 56.7 | - | 52.5† | 50.6† | 58.8 |
| Chinese | ||||||||
| C‑EVAL | 5-shot | 87.2 | 89.6 | 87.9 | 86.9 | 90.0† | 91.0† | 92.5 |
| CMMLU | 5-shot | 86.9 | 88.9 | 87.4 | - | 88.8† | 88.9† | 90.9 |
| C‑SimpleQA | 5-shot | 58.1 | 63.2 | 61.5 | 70.1 | 70.9† | 68.0† | 77.6 |
temperature=0.6, top_p=0.95temperature=1.0, top_p=0.95.Step 3.5 Flash is built on a Sparse Mixture-of-Experts (MoE) transformer architecture, optimized for high throughput and low VRAM usage during inference.
| Component | Specification |
|---|---|
| Backbone | 45-layer Transformer (4,096 hidden dim) |
| **Cont |
From the published model card. Full card on the HuggingFace links in the sidebar.
Using it via the API
Once AxForge deploys step-3-5-flash-midtrain for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (step-3-5-flash-midtrain below is illustrative; you get the exact model name on deployment.)
$ curl -sS https://api.axforge.ai/v1/chat/completions \
-H "Authorization: Bearer $AXFORGE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"step-3-5-flash-midtrain","messages":[{"role":"user","content":"Hello"}]}'
Create an account — your API key is available in the console. 5M tokens/month currently included with every new account at launch.