Developer
StepFun

Step-3.7-Flash

StepFun's open-weight vision-language flagship, released May 29, 2026. A 198B-parameter MoE that activates ~11B per token, serves at over 400 tokens/second under Apache 2.0, and trades benchmark prestige for throughput and price.

Released
May 29, 2026
Type
Multimodal Language Model
Context window
256K tokens
Pricing
$0.20 / $1.15 per Mtok
License
Apache 2.0
🌐 Official Website📚 Documentation💰 Pricing & API
On this page

Overview

Step-3.7-Flash is StepFun's open-weight vision-language flagship, released on May 29, 2026. StepFun's model card describes it as "a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model that combines a 196B-parameter language backbone with a 1.8B-parameter vision encoder for native image understanding." It "activates approximately 11B parameters per token and delivers a throughput of up to 400 tokens per second," supports "a 256k context window," and "offers three selectable reasoning levels (low, medium, and high) so developers can easily balance speed, cost, and cognitive depth."

StepFun (阶跃星辰; legally 上海阶跃星辰智能科技有限公司) was founded on April 6, 2023 by Jiang Daxin, Zhu Yibo and Jiao Binxing. The positioning of this release is stated plainly on its model page — "The new frontier is agent efficiency" — and in its GitHub repository description: "A high-efficiency Flash model for real-world agents." StepFun says it was built "for developers who need to scale agentic workflows that combine perception, search, and reasoning."

That framing is worth taking literally, because the independent numbers reward throughput rather than intelligence. Artificial Analysis clocks Step-3.7-Flash at 406.7 output tokens per second — first of 93 models in its comparison set, exceeding the vendor's "up to 400 tokens per second" claim. On the same site it scores 30 on Intelligence Index v4.1, ranking #32 of 93 open-weights models. Against the current Chinese flagship cohort that is a wide gap: Artificial Analysis's leaderboard places MiniMax-M3 at 44, Kimi K2.6 at 44, MiMo-V2.5-Pro at 42, Qwen3.7-Max at 46, and GLM-5.2 at 51. The honest summary is volume without prestige — a model chosen for throughput and unit economics, not for topping leaderboards.

A note on naming

Three claims circulating about this model are false, and one is a near-miss worth untangling.

  • "Step-R" does not exist. No such model appears in StepFun's Hugging Face organisation or model listings. The confusion likely stems from Step-R1-V-Mini, a genuine StepFun multimodal reasoning model from April 2025 — unrelated to the 3.7 generation.
  • "Step-4" does not exist. StepFun has never released a model under that name. The lineage runs Step-1 → Step-2 → Step 3 (July 2025) → Step 3.5 Flash (early 2026) → Step-3.7-Flash.
  • There is no larger, non-Flash Step-3.7. stepfun-ai/Step-3.7 does not resolve as a public repository. Flash is the release. The only siblings are quantisations of the same checkpoint: -FP8, -NVFP4, and -GGUF.

Capabilities

  • Native visual understanding: A 1.8B perception-encoder ViT (728px images, 14px patches, 47 layers) is fused with the language backbone. StepFun claims the model "accurately processes dense visual interfaces, such as UI wireframes, application GUIs, and data charts, to map them into structured code."
  • Three-tier reasoning with mandatory chain-of-thought: low, medium, and high. OpenRouter's API flags reasoning as mandatory: true with a default effort of medium — this model always thinks before answering.
  • Tool orchestration: StepFun reports 67.1 on ClawEval-1.1, claiming it "significantly outperforms the next closest competitor at 59.8," and describes "high resistance to adversarial traps and strict adherence to system policies during multi-turn orchestration."
  • Search-augmented perception: "When it encounters an incomplete visual asset, it can independently identify missing data and execute lookups to verify context before returning a factually verified conclusion."
  • Throughput as a design goal: Up to 400 tokens/second, independently corroborated by Artificial Analysis at 406.7 tok/s.
  • Speculative decoding built in: The config declares three Multi-Token Prediction layers, and StepFun's vLLM recipe enables them with --speculative_config '{"method": "mtp", "num_speculative_tokens": 3}'.

Technical Specifications

  • Hugging Face repository: stepfun-ai/Step-3.7-Flash
  • Total parameters (vendor headline): 198B — 196B language backbone + 1.8B ViT
  • Total parameters (safetensors index): 201,365,316,160
  • Active parameters: ~11B per token
  • Architecture: Mixture-of-Experts transformer, 288 experts, top-8 routing, sigmoid router
  • Layers: 45, of which layers 3–44 are MoE; hidden size 4096; GQA attention (64 heads, 8 groups); 512-token sliding window
  • MTP layers: 3 (num_nextn_predict_layers)
  • Vision encoder: perception_encoder, 728×728 input, patch size 14, width 1536, 47 layers
  • Context window: 256K tokens (max_position_embeddings: 262144), reached by llama3 RoPE scaling at factor 2.0 from a native 131,072
  • Modalities: OpenRouter lists text+image+video->text; the model card itself describes only image understanding
  • License: Apache 2.0
  • model_type: step3p7; the text backbone still reports step3p5, confirming it is the Step-3.5-Flash lineage

The 198B headline and the 201.4B safetensors count do not agree. The model card offers no reconciliation, and this page does not guess at one. StepFun publishes no knowledge cutoff and no training-token count, so this page does not state either.

Use Cases

  • High-volume agent fleets: The stated design target — "operating concurrent coding agents in high-throughput pipelines," where per-token cost and tokens/second dominate the economics.
  • Document and report parsing: StepFun cites "parsing massive financial reports in one pass," which the 256K context window supports.
  • Screenshot-to-code and GUI automation: The vision encoder is aimed squarely at wireframes, GUIs and charts.
  • Multi-step search loops: "running multi-step search loops with cross-source verification," per the model card.
  • Latency-sensitive interactive products: 406.7 tok/s is the fastest measured in Artificial Analysis's 93-model cohort; reasoning_effort: low trims the mandatory thinking budget.
  • Cost-constrained self-hosting: 11B active parameters run on a NVIDIA DGX Station, an AMD Ryzen AI Max+ 395 system, or a Mac Studio / MacBook Pro with ≥128GB unified memory, per StepFun's deployment notes.

Performance / Benchmarks

Vendor-reported. Step-3.7-Flash's own scores are quoted from the text of StepFun's Hugging Face model card; the competitor column is read off the benchmark table on StepFun's model page, whose comparison set includes DeepSeek V4 Flash, Gemini 3.5 Flash, GPT-5.5, Claude Opus 4.7, Kimi K2.6 and GLM-5.1.

BenchmarkStep-3.7-FlashStepFun's cited leader
ClawEval-1.167.1next closest 59.8
SimpleVQA (Search)79.2GPT-5.5: 79.1
V* (with Python)95.3Kimi K2.6: 96.9
SWE-Bench Pro56.3Claude Opus 4.7: 64.3
Terminal-Bench 2.159.5GPT-5.5: 82.7
Toolathlon49.5Claude Opus 4.7: 65.4
HLE w. Tool48.1not stated
GDPVal-AA45.8not stated

StepFun concedes the weak spots itself: "evaluations like Terminal-Bench 2.1 (59.5) and GDPVal-AA (45.8) show clear areas for future optimization compared to the absolute peak of the cohort."

Third-party placements

Artificial Analysis (Intelligence Index v4.1, retrieved 2026-07-08): Intelligence 30, output speed 406.7 tok/s, #1 of 93. Intelligence rank #32 of 93 open-weights models, which Artificial Analysis describes as "above average among other open weight models of similar size" against a median of 25. OpenRouter's model API mirrors the same Artificial Analysis figures at finer precision — intelligence 29.7, coding index 37.3, agentic index 21.5 — the latter two are not shown on Artificial Analysis's own model page.

Design Arena (per-category ranks, as embedded in OpenRouter's model API): ASCII art #20 (Elo 1210), dataviz #43 (1212), website #44 (1225), code categories #45 (1217), UI component #45 (1212), game dev #46 (1205), 3D #48 (1197), SVG #50 (1125).

arena.ai: not present. Its leaderboard changelog logs step-3.5-flash being "added to the Text leaderboard" on February 10, 2026, but carries no entry for Step-3.7-Flash.

Note that Artificial Analysis and arena.ai are different organisations with different scales; their numbers are not cross-comparable.

Limitations

  • Intelligence is the trade: 30 on Intelligence Index v4.1 sits far below every current Chinese open-weight flagship — GLM-5.2 (51), Qwen3.7-Max (46), MiniMax-M3 and Kimi K2.6 (44), MiMo-V2.5-Pro (42). Buying this model for reasoning quality is a mistake.
  • Verbose, which erodes the price advantage: Artificial Analysis measured 260M output tokens across its evaluation suite versus a 92M median for comparable models. At $1.15 per million output tokens, roughly 2.8× the median verbosity substantially narrows the apparent cost gap.
  • Reasoning cannot be switched off: OpenRouter reports reasoning.mandatory: true. There is no no_think mode, so even trivial turns pay a thinking-token toll.
  • Benchmarks are vendor-selected: The table above is StepFun's, run against StepFun's chosen competitor set. Independent replication is thin. Treat the ClawEval-1.1 and SimpleVQA leads as upper bounds.
  • Parameter count is internally inconsistent: 198B claimed, 201.4B in the safetensors index, unreconciled.
  • Two incompatible API regions: Keys issued on platform.stepfun.ai (global) do not work against api.stepfun.com (China), and vice versa — StepFun warns requests "will be rejected as unauthorized."
  • Little arena presence: Absent from arena.ai's leaderboard, and seven of its eight Design Arena category ranks sit in the 40s–50s (ASCII art, at #20, is the lone outlier). There is no large-scale human-preference signal for this model.
  • No published knowledge cutoff: Ground time-sensitive queries with retrieval or tool use.

Pricing & Access

Self-hosting is free. The weights are Apache 2.0 and published at stepfun-ai/Step-3.7-Flash, with FP8, NVFP4 and GGUF quantisations in the same organisation.

Hosted access. StepFun's model card publishes a rate card directly, in dollars per million tokens, and OpenRouter's API reports identical figures:

Token typePrice per 1M tokens
Input (cache miss)$0.20
Input (cache hit)$0.04
Output$1.15

Available on the StepFun Open Platform — platform.stepfun.ai (global) and platform.stepfun.com (China) — plus OpenRouter and NVIDIA NIM. StepFun states it "is also partnering with DeepInfra, Fireworks AI, and Modal to expand availability soon"; the model card still described those three as forthcoming as of July 8, 2026.

Serving: vLLM (vllm/vllm-openai:stepfun37), SGLang (lmsysorg/sglang:dev-step-3.7-flash), Hugging Face Transformers, and llama.cpp. The reasoning and tool-call parsers are both named step3p5.

Adoption signal: 146,515 downloads in the trailing 30 days and 408 likes on the base repository, with a further 157,601 on the NVFP4 quantisation — which, unusually, out-downloads the BF16 original. Across all 51 repositories in the stepfun-ai organisation, downloads total 1,484,107 for the period; the single most-downloaded repo is not Step-3.7-Flash but stepfun-ai/Step3-VL-10B at 325,897.

Ecosystem & Tools

  • stepfun-ai/Step-3.7-Flash on Hugging Face — Apache 2.0 weights, benchmark table, deployment recipes
  • Step-3.7-Flash on GitHub — "A high-efficiency Flash model for real-world agents"
  • StepFun Open Platform — hosted API, global region
  • Step-3.7-Flash on OpenRouter — third-party hosted access
  • StepAudio 2.5 Realtime — StepFun's end-to-end speech model (May 2026), "an end-to-end real-time speech large language model" with "fully customizable persona capabilities"; technical report. API-only — no open weights published
  • ACE-Step 1.5 — music generation co-developed by ACE Studio and StepFun; the Hugging Face repo ACE-Step/Ace-Step1.5 carries an MIT license tag (not Apache 2.0), was created January 23, 2026, and has 45,522 downloads / 791 likes
  • stepfun-ai/Step-Audio-EditX and the Step-Audio-R1 line — speech editing and audio reasoning models, evidence of an unusually broad multimodal portfolio for a company of StepFun's size

Community & Resources

Frequently Asked Questions

May 29, 2026, per StepFun's own model page. OpenRouter's canonical slug for the model is stepfun/step-3.7-flash-20260528, which reflects the UTC date of its listing — StepFun is a Beijing-time company, and the two dates are the same moment. The Hugging Face repository was created earlier, on May 23, 2026, before the public announcement.
Yes, by the usual standard. The weights at stepfun-ai/Step-3.7-Flash carry a license:apache-2.0 tag and the GitHub repository is Apache 2.0. There are no field-of-use or geographic restrictions. Training data is not released, so this is open weights under an OSI-approved license rather than a fully reproducible model.
StepFun's model card publishes a USD rate card: $0.20 per million input tokens (cache miss), $0.04 per million on a cache hit, and $1.15 per million output tokens. OpenRouter's API reports the identical figures. Self-hosting the Apache 2.0 weights is free.
No. Neither name appears in StepFun's Hugging Face organisation, its model family listings, or its Wikipedia entry. The likely source of "Step-R" confusion is Step-R1-V-Mini, a real multimodal reasoning model StepFun shipped in April 2025 — a different and much older model. stepfun-ai/Step-4 and stepfun-ai/Step-R are not publicly resolvable repositories.
No. Flash is the entire 3.7 release. stepfun-ai/Step-3.7 does not resolve publicly, and no StepFun model listing includes a non-Flash 3.7 variant. The only sibling repositories are quantisations of the same model: -FP8, -NVFP4, and -GGUF.
StepFun's model card states "a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model that combines a 196B-parameter language backbone with a 1.8B-parameter vision encoder," activating "approximately 11B parameters per token." The Hugging Face safetensors index reports a higher total of 201,365,316,160 parameters. StepFun does not reconcile the two figures.
Mostly fast. Artificial Analysis measures 406.7 output tokens/second — ranked #1 of 93 models in its comparison set — but scores it 30 on Intelligence Index v4.1, versus 44 for MiniMax-M3, 46 for Qwen3.7-Max, and 51 for GLM-5.2. Artificial Analysis does note the score is "above average among other open weight models of similar size," whose median is 25.
Because Artificial Analysis has revised its Intelligence Index and older versions scored models differently. The figure on this page — 30, or 29.7 as mirrored by OpenRouter's API — is explicitly labelled Intelligence Index v4.1 on Artificial Analysis's model page. Higher numbers from earlier index versions are not comparable and should not be quoted alongside v4.1 scores.
No. arena.ai's leaderboard changelog records step-3.5-flash being "added to the Text leaderboard" on February 10, 2026, and Step 3 being added to "the Text and Vision leaderboards" on August 22, 2025. There is no changelog entry for Step-3.7-Flash as of July 8, 2026.
An unusually broad multimodal portfolio for a company of its size: StepAudio 2.5 Realtime (an end-to-end speech model, May 2026), Step-Audio-EditX, the Step-Audio-R1 reasoning line, the NextStep image generators, and ACE-Step 1.5 — an MIT-licensed music generation model co-developed with ACE Studio.

Related Models

MiniMax-M3

MiniMax's native-multimodal flagship, released June 1, 2026. A 428B-parameter MoE with ~23B active parameters, a 1M-token context window built on MiniMax Sparse Attention, and open weights under a restricted community license — not the MIT terms of the M2 generation.

MiMo-V2.5-Pro

Xiaomi's open-weight flagship, released April 22, 2026, with weights published April 27. A 1.02-trillion-parameter Mixture-of-Experts model with 42B active parameters, a 1M-token context window, and an MIT license.

GLM-5.2

Zhipu AI's open-weight flagship: 753B parameters, a sparse GLM-MoE-DSA attention stack with IndexShare, a genuinely usable 1M token context window, and an MIT license.

DeepSeek V4

DeepSeek's open-weight MoE flagship, shipped as V4-Pro (1.6T total / 49B active) and V4-Flash (284B / 13B), with a hybrid sparse attention stack, a 1M token context window, and an MIT license.

Kimi K2.6

Moonshot AI's flagship open-weight model, released April 20, 2026. A natively multimodal 1T-parameter MoE with 32B active parameters, a MoonViT vision encoder, a 256K context window, and a Modified MIT license.

Tencent Hy3

Tencent's open-weight flagship, released July 6, 2026. A 295B-parameter MoE with 21B active parameters, hybrid fast-and-slow thinking, a 256K context window, and an Apache 2.0 license.

Explore More Models

Discover other AI models and compare their capabilities.

Browse Models← Back to Models