StepFun Releases Step 3.7 Flash: A 198B MoE Vision-Language Model for Coding Agents and Search Workflows
StepEnjoyable right this moment launched Step 3.7 Flash, a multimodal Mixture-of-Experts mannequin concentrating on agentic use circumstances. It provides native imaginative and prescient enter and improved tool-use reliability over Step 3.5 Flash.
What is Step 3.7 Flash?
Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language mannequin. It pairs a 196B-parameter language spine with a 1.8B-parameter imaginative and prescient encoder (ViT) for native picture understanding.
The mannequin prompts roughly 11B parameters per token throughout inference. In MoE architectures, solely a subset of “skilled” sub-networks fires per ahead cross — not the complete community. This retains inference compute nearer to an 11B dense mannequin whereas sustaining a 198B complete parameter funds.
Key specs:
| Spec | Value |
|---|---|
| Total parameters | 198B (196B language + 1.8B ViT) |
| Active parameters per token | ~11B |
| Context window | 256k tokens |
| Throughput | Up to 400 tokens/sec |
| Reasoning ranges | Low, medium, excessive |
| License | Apache 2.0 |
Architecture Notes
The imaginative and prescient encoder runs as a separate 1.8B ViT module. It injects picture representations into the language spine’s context. Step 3.5 Flash had no multimodal assist; it is a new addition in 3.7.
Three selectable reasoning depths — low, medium, and excessive — let builders commerce latency for reasoning depth. Low is quicker and cheaper; excessive applies extra computation per response.
Agentic Coding Performance
On SWE-Bench Pro, Step 3.7 Flash scores 56.26%, up from Step 3.5 Flash’s 51.3% — a achieve of roughly 5 share factors. On Terminal-Bench 2.1, it scores 59.55%, up from 53.37%.
On SWE-MTLG (a multi-task long-generation coding benchmark), it scores 72.42%.
Cross-harness consistency on StepEnjoyable’s inside Step-SWE-Bench:
| Scaffold | Step 3.7 Flash | Step 3.5 Flash |
|---|---|---|
| Hermes Agent | 67.5% | 60.0% |
| OpenClaw | 67.0% | 47.0% |
| KiloCode | 67.5% | 59.0% |
| RooCode | 64.5% | 43.0% |
| Claude Code | 71.5% | 73.0% |
| OpenCode | 64.5% | 57.0% |
Step 3.5 Flash ranged from 43% to 73% throughout harnesses. Step 3.7 Flash ranges from 64.5% to 71.5%. In manufacturing, coding brokers usually run inside heterogeneous scaffolds — every with its personal prompting conventions and device schemas. Narrower per-harness variance means extra predictable conduct throughout totally different setups.
Advisor Mode
Step 3.7 Flash helps Advisor Mode, StepEnjoyable’s implementation of the advisor technique described by Anthropic. The mannequin runs the agentic loop end-to-end — calling instruments, studying outcomes, iterating — and escalates to a bigger advisor mannequin solely at particular inflection factors, corresponding to planning or recovering from repeated failures. Most of the run stays at executor price.
With Advisor Mode enabled on SWE-Bench Verified, StepEnjoyable reviews Step 3.7 Flash reaches 97% of Claude Opus 4.6’s coding efficiency at roughly one-ninth the per-task price ($0.19 vs. $1.76 per activity). These are StepEnjoyable’s inside figures.
Multimodal Capabilities
Step 3.7 Flash helps two visible device pathways:
Visual Search Tool — For recognition duties the place the mannequin’s parametric data is inadequate (long-tail entities, lately emerged ideas), it invokes a visible search device to retrieve and confirm. On SimpleVQA (with Search), it scores 79.16%, akin to GPT 5.5 (79.11%) and above Kimi K2.6 (78.24%) and GLM 5V Turbo (78.20%).
Python Tool — For fine-grained visible duties (high-resolution photos, visible probing, bounding-box evaluation), it makes use of a code interface to crop, zoom, and draw pixels or bounding containers. On V (a self-tested rating with Python), it scores 95.29%. On HR-Bench 4K and HR-Bench 8K, it scores 89.13% and 86.34% respectively.
StepEnjoyable notes an noticed conduct throughout testing: the mannequin mixed visible instruments with non-visual instruments with out being explicitly educated to take action. For instance, after producing frontend code, it used the GUI to render and examine the outcome earlier than iterating. StepEnjoyable describes this as emergent compositional device use.
On Android Daily (long-horizon cellphone UI activity completion), Step 3.7 Flash scores 61.87%, forward of Kimi K2.6 (53.36%) and GLM 5V Turbo (51.68%). Gemini 3 Flash (63.21%) leads this benchmark.
Search and Research Benchmarks
StepEnjoyable targeted this mannequin’s search design on planning, proof filtering, and synthesis — integrating search as a part of the reasoning loop relatively than a separate add-on.
| Benchmark | Step 3.7 Flash | Notable comparability |
|---|---|---|
| HLE with Tools (acc) | 47.20% | DeepSeek V4 Flash: 45.10% |
| BrowseComp (acc) | 75.82% | Claude Opus 4.7: 79.30% |
| DeepSearchQA (F1) | 92.82% | Kimi K2.6: 92.50% |
| ResearchRubrics (rating) | 71.68% | GPT 5.5: 61.50% |
Note: The HLE with Tools rating of 47.20% compares to Step 3.5 Flash’s text-only rating of 35.68%. Step 3.5 Flash didn’t assist tool-augmented analysis on HLE.
General Agent Benchmarks
| Benchmark | Step 3.7 Flash | Description |
|---|---|---|
| Toolathlon | 49.51% | Multi-tool coordination |
| ClawEval-1.1 | 67.07% | Daily autonomous activity execution in practical environments |
| GDPval (44 occupations) | 45.8% | General skilled activity execution |
| Tau2-bench Telecom | >98% | Across totally different reasoning issue tiers |
On ClawEval-1.1, Step 3.7 Flash (67.07%) leads DeepSeek V4 Flash (57.80%) and DeepSeek V4 Pro (59.80%) among the many in contrast fashions.
Long-Context Performance
On AA-LCR (a long-context retrieval benchmark, avg@16/acc), Step 3.7 Flash scores 63.94%. This is akin to DeepSeek V4 Flash (63.70%) and DeepSeek V4 Pro (66.30%).
Pricing
| Token Type | Price |
|---|---|
| Input (cache miss) | $0.20 / M tokens |
| Input (cache hit) | $0.04 / M tokens |
| Output | $1.15 / M tokens |
Marktechpost’s Visual Explainer
Key Takeaways
- Step 3.7 Flash is a 198B sparse MoE mannequin with 11B lively params and a 256k context window.
- Native multimodal assist (photos, GUIs, paperwork) is new — Step 3.5 Flash was text-only.
- Advisor Mode reaches 97% of Claude Opus 4.6's SWE-Bench Verified efficiency at $0.19 per activity vs. $1.76.
- Cross-harness coding variance narrowed from a 43–73% vary (3.5 Flash) to 64.5–71.5% (3.7 Flash).
- Released below Apache 2.0 with BF16, FP8, NVFP4, and GGUF weights on Hugging Face.
Where (Inferences) to Run Step 3.7 Flash
Live
Live
Live
Day-0
Weights
Repo
Check out the Model Weights, Repo and Technical Details. Also, be happy to observe us on Twitter and don’t neglect to affix our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to companion with us for selling your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar and so forth.? Connect with us
The publish StepFun Releases Step 3.7 Flash: A 198B MoE Vision-Language Model for Coding Agents and Search Workflows appeared first on MarkTechPost.
