Overview
Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts vision-language model released by StepFun AI on May 29, 2026. Despite its massive total parameter count, it activates only ~11B parameters per token through expert routing (8 out of 288 experts), delivering near-frontier intelligence at inference costs comparable to small dense models. The model combines a 196B language backbone with a 1.8B native vision encoder, supports a 256K context window, and achieves up to 400 tokens/second throughput on cloud APIs.
The model's primary differentiation lies in agentic workflow reliability rather than absolute intelligence ceiling. It leads the ClawEval-1.1 benchmark at 67.1 (a 23.5-point jump over its predecessor Step 3.5 Flash) and places second on SWE-Bench PRO at 56.3. Its most innovative feature is Advisor Mode, a two-tier execution strategy where Step 3.7 Flash handles routine tool-calling and iteration while escalating to a larger frontier model only at genuine decision points, achieving approximately 97% of Claude Opus 4.6's coding performance at roughly one-ninth the per-task cost ($0.19 vs $1.76).
The model is fully open-weight under Apache 2.0 and ships with quantized checkpoints (BF16, FP8, NVFP4, GGUF) across HuggingFace, enabling local deployment on 128GB+ unified memory devices like NVIDIA DGX Spark and Mac Studio. It is available via StepFun's own platform, OpenRouter, NVIDIA NIM, and forthcoming partnerships with DeepInfra, Fireworks AI, and Modal. Where it falls short is on pure reasoning benchmarks (HLE w/ tools at 47.2% vs Claude Opus 4.7's 54.7%), terminal-heavy workflows (Terminal-Bench 2.1 at 59.5 vs GPT 5.5's 82.7), and general professional tasks (GDPVal-AA at 45.8% vs frontier models at 63%).
Benchmarks & Performance
## Benchmark Performance
### Agentic & Tool Use
| Benchmark | Step 3.7 Flash | Claude 4 Opus | GPT 5.5 | DeepSeek V4 Flash | Gemini 3.5 Flash |
|---|---|---|---|---|---|
| ClawEval-1.1 | **67.1** (#1) | 59.8 | 60.3 | 43.6 | 57.8 |
| Toolathlon | 49.5 | N/A | N/A | N/A | N/A |
| HLE w. Tools | 47.2 | N/A | 52.2 | N/A | N/A |
| BrowseComp | 75.82% | 79.3%* | N/A | N/A | N/A |
### Coding
| Benchmark | Step 3.7 Flash | Claude 4 Opus | GPT 5.5 | DeepSeek V4 Flash | Gemini 3.5 Flash |
|---|---|---|---|---|---|
| SWE-Bench PRO | 56.3 (#2) | **64.3** (#1) | 58.6 | 55.6 | 55.1 |
| SWE-Bench Verified (Advisor Mode) | 76.3% | N/A | N/A | N/A | N/A |
| Terminal-Bench 2.1 | 59.5 | N/A | 82.7* | N/A | 76.2* |
### Multimodal & Visual
| Benchmark | Step 3.7 Flash | Notes |
|---|---|---|
| SimpleVQA (Search) | **79.2** (#1) | Beats GPT 5.5 (79.1) |
| V* (Python) | 95.3 | Competitive with Kimi K2.6 (96.9) and Gemini 3 Flash (96.3) |
| AA-MMMU-Pro | 75.3% | N/A |
### Knowledge & Reasoning
| Benchmark | Step 3.7 Flash | Notes |
|---|---|---|
| DeepSearchQA (F1) | 92.8% | +11pp over Step 3.5 Flash |
| GDPVal-AA | 45.8% | vs GPT 5.5 & Claude Opus 4.7 at 63% |
| Toolathlon | 49.5% | N/A |
### NVFP4 + Speculative Decoding Throughput
| Concurrency | NVFP4 + MTP | NVFP4 (no MTP) | Speedup |
|---|---|---|---|
| 8 | 1,309 tok/s | 1,155 tok/s | 1.13× |
| 32 | 4,391 tok/s | 3,480 tok/s | 1.26× |
| 64 | **8,229 tok/s** | 5,667 tok/s | **1.45×** |
*Benchmark data sourced from StepFun's official GitHub repository and model page. Some competitor scores are from StepFun's own evaluation runs. Verified Advisor Mode cost: $0.19/task (vs Claude Opus 4.6 at $1.76/task) per Flowtivity's independent testing.
Detailed Comparison
## Head-to-Head Comparisons
### Step 3.7 Flash vs Claude Opus 4.6
| Dimension | Step 3.7 Flash | Claude Opus 4.6 |
|---|---|---|
| Input Price | $0.20/M | $5.00/M |
| Output Price | $1.15/M | $25.00/M |
| SWE-Bench PRO | 56.3 | 64.3 |
| ClawEval-1.1 | 67.1 | 59.8 |
| Context Window | 256K | 200K |
| Open Weight | Yes (Apache 2.0) | No |
**Verdict:** Claude Opus wins on pure coding and reasoning depth. Step 3.7 Flash wins on agent reliability, cost efficiency (25× cheaper on input), and deployability. For production agent workflows where tool orchestration matters more than single-turn intelligence, Step 3.7 Flash at 1/9th the cost is compelling.
### Step 3.7 Flash vs DeepSeek V4 Flash
| Dimension | Step 3.7 Flash | DeepSeek V4 Flash |
|---|---|---|
| Input Price | $0.20/M | $0.22/M |
| Output Price | $1.15/M | $0.87/M |
| SWE-Bench PRO | 56.3 | 55.6 |
| ClawEval-1.1 | 67.1 | 43.6 |
| Parameters (total) | 198B | 284B |
| Context Window | 256K | 1M+ |
| Vision | Native (1.8B ViT) | Separate model |
**Verdict:** Step 3.7 Flash is cheaper on input, smaller in total parameters, yet decisively better on agent benchmarks (ClawEval gap of 23.5 points). DeepSeek V4 Flash offers a much larger context window. For agentic workloads, Step 3.7 Flash is the stronger choice.
### Step 3.7 Flash vs Gemini 3.5 Flash
| Dimension | Step 3.7 Flash | Gemini 3.5 Flash |
|---|---|---|
| Input Price | $0.20/M | $0.15/M |
| Output Price | $1.15/M | $0.60/M |
| SWE-Bench PRO | 56.3 | 55.1 |
| ClawEval-1.1 | 67.1 | 57.8 |
| Context Window | 256K | 1M |
| Open Weight | Yes | No |
**Verdict:** Gemini 3.5 Flash is cheaper per token and offers a 4× larger context window. Step 3.7 Flash leads on both SWE-Bench PRO (+1.2pp) and ClawEval (+9.3pp) and is fully open-weight. For teams needing agent reliability with self-hosting capability, Step 3.7 Flash is preferable. For long-context cloud-native work, Gemini may be better.
Community Feedback
The developer community response to Step 3.7 Flash has been notably enthusiastic, particularly around practical deployment rather than benchmark performance alone. On X/Twitter, multiple users shared working deployment configs for running the 198B model on NVIDIA DGX Spark desktops and Mac Studio devices, with one widely-shared post showing a DGX Spark on a regular desk running the model with detailed llama.cpp configuration.
The Flowtivity team reported a 100% tool call success rate in their agentic testing suite and switched their daily driver from MiniMax M2.7 to Step 3.7 Flash. They noted the ClawEval-1.1 improvement from 43.6 (Step 3.5) to 67.1 (Step 3.7) as a 'generational leap in agent capability.'
Community discussions on Hacker News have been more measured, with some users noting that StepFun's previous model (Step 3.5 Flash) received mixed reviews and expressing caution about self-reported benchmark numbers. The LavX News analysis highlighted that many competitor scores in StepFun's benchmark tables come from their own re-runs rather than standardized evaluations, which introduces potential bias.
Developer adoption is accelerating across agent frameworks. Step 3.7 Flash is explicitly supported by OpenClaw, Hermes Agent, Kilo Code, and compatible with any OpenAI-compatible endpoint including Aider and Continue. The day-one availability of four quantization formats (BF16, FP8, NVFP4, GGUF) was widely praised as a model for how open-weight releases should be handled.
The Advisor Mode concept has generated significant discussion. While the $0.19/task vs $1.76/task comparison is frequently cited, some reviewers (LavX News, WhyOps) caution that complex real-world pipelines may trigger more advisor escalations than benchmark scenarios suggest, potentially eroding the 9× cost advantage.
Use Cases
### 1. Production Agentic Coding Pipelines
Step 3.7 Flash excels as the executor in multi-step coding agents. Its ClawEval-1.1 score of 67.1 indicates high reliability in maintaining trajectory integrity across long tool-calling chains. Use it with frameworks like OpenClaw or Kilo Code for automated bug fixing, code review, and repository-scale refactoring. In Advisor Mode, pair it with a frontier model (Claude Opus, GPT 5.5) that only activates at planning or failure-recovery inflection points, achieving ~97% of frontier quality at ~1/9th the cost. **Choose over alternatives when:** cost-per-task matters and your agent makes dozens of sequential tool calls per task.
### 2. Multimodal Document Processing & Visual QA
With its native 1.8B vision encoder and #1 ranking on SimpleVQA (Search) at 79.2, Step 3.7 Flash is well-suited for parsing dense financial reports, extracting data from charts and UI wireframes, and performing cross-source verification with visual grounding. It can process screenshots, identify missing data independently, and execute lookups to verify context. **Choose over alternatives when:** your workflow requires interleaving image understanding with search-augmented reasoning, especially for document-heavy industries like finance or legal.
### 3. High-Throughput Batch Processing with NVFP4
For teams running large-scale agent swarms, the NVFP4 quantization with MTP speculative decoding delivers up to 8,229 tok/s at concurrency 64 on GB200 hardware (1.45× speedup over non-MTP). This makes Step 3.7 Flash practical for processing thousands of concurrent agent tasks — batch code generation, automated testing pipelines, or high-volume document analysis. Quality remains statistically equivalent to NVFP4 without MTP (77.81% vs 78.41% on GPQA Diamond). **Choose over alternatives when:** you need maximum throughput on NVIDIA hardware and can accept NVFP4 precision.
### 4. Self-Hosted Enterprise Agent Deployment
With Apache 2.0 licensing and deployment support on 128GB+ devices (Mac Studio, DGX Spark, AMD Ryzen AI Max+ 395), Step 3.7 Flash enables fully air-gapped agent deployments where data never leaves corporate infrastructure. The Flowtivity team demonstrated production deployment on DGX Spark with ~27 tok/s generation speed, replacing their previous daily driver model. **Choose over alternatives when:** data sensitivity or regulatory requirements prohibit cloud API usage and you need a capable agent model that runs on workstation-class hardware.
Latest News
### Release Timeline (May–June 2026)
- **May 29, 2026:** StepFun officially released Step 3.7 Flash with simultaneous availability of BF16, FP8, NVFP4, and GGUF checkpoints on HuggingFace.
- **Day 1 availability:** StepFun Platform (Global: platform.stepfun.ai, China: platform.stepfun.com), OpenRouter, and NVIDIA NIM.
- **Forthcoming partnerships:** DeepInfra, Fireworks AI, and Modal announced for expanded cloud availability.
### NVFP4 + MTP Speculative Decoding (May 29, 2026)
The NVFP4 checkpoint includes MTP (Multi-Token Prediction) draft layers supporting vLLM speculative decoding with up to 1.45× throughput improvement at high concurrency on GB200 TP=4 setups. Quality remains statistically equivalent to non-MTP NVFP4.
### Ecosystem Integration (May 29, 2026)
- **Inference:** vLLM (prebuilt Docker image `vllm/vllm-openai:stepfun37`), SGLang (Docker image `lmsysorg/sglang:dev-step-3.7-flash`), HuggingFace Transformers 5.0+, and llama.cpp (custom `step3.7` branch).
- **Development:** NVIDIA NeMo ecosystem support including AutoModel, Megatron Core, and Megatron Bridge.
- **Agent platforms:** Explicit support for OpenClaw, Hermes Agent, Kilo Code.
### Open-Source Release (May 29, 2026)
Full model weights released under Apache 2.0 on both HuggingFace and GitHub (stepfun-ai/Step-3.7-Flash). The GitHub repository includes deployment instructions for cloud, data center, and local environments. Custom llama.cpp fork maintained at github.com/stepfun-ai/llama.cpp (branch `step3.7`).