{"id":1007,"date":"2026-05-30T05:25:05","date_gmt":"2026-05-29T21:25:05","guid":{"rendered":"https:\/\/connectword.dpdns.org\/?p=1007"},"modified":"2026-05-30T05:25:05","modified_gmt":"2026-05-29T21:25:05","slug":"stepfun-releases-step-3-7-flash-a-198b-moe-vision-language-model-for-coding-agents-and-search-workflows","status":"publish","type":"post","link":"https:\/\/connectword.dpdns.org\/?p=1007","title":{"rendered":"StepFun Releases Step 3.7 Flash: A 198B MoE Vision-Language Model for Coding Agents and Search Workflows"},"content":{"rendered":"

StepFun today released Step 3.7 Flash<\/a>, a multimodal Mixture-of-Experts model targeting agentic use cases. It adds native vision input and improved tool-use reliability over Step 3.5 Flash.<\/p>\n

What is Step 3.7 Flash?<\/strong><\/h2>\n

Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model<\/strong>. It pairs a 196B-parameter language backbone<\/strong> with a 1.8B-parameter vision encoder (ViT)<\/strong> for native image understanding.<\/p>\n

The model activates approximately 11B parameters per token<\/strong> during inference. In MoE architectures, only a subset of \u201cexpert\u201d sub-networks fires per forward pass \u2014 not the full network. This keeps inference compute closer to an 11B dense model while maintaining a 198B total parameter budget.<\/p>\n

Key specs:<\/strong><\/p>\n

\n\n\n\n\n\n\n\n\n\n\n
Spec<\/th>\nValue<\/th>\n<\/tr>\n<\/thead>\n
Total parameters<\/td>\n198B (196B language + 1.8B ViT)<\/td>\n<\/tr>\n
Active parameters per token<\/td>\n~11B<\/td>\n<\/tr>\n
Context window<\/td>\n256k tokens<\/td>\n<\/tr>\n
Throughput<\/td>\nUp to 400 tokens\/sec<\/td>\n<\/tr>\n
Reasoning levels<\/td>\nLow, medium, high<\/td>\n<\/tr>\n
License<\/td>\nApache 2.0<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n
\n

Architecture Notes<\/strong><\/h2>\n

The vision encoder runs as a separate 1.8B ViT module. It injects image representations into the language backbone\u2019s context. Step 3.5 Flash had no multimodal support; this is a new addition in 3.7.<\/p>\n

Three selectable reasoning depths \u2014 low, medium, and high \u2014 let developers trade latency for reasoning depth. Low is faster and cheaper; high applies more computation per response.<\/p>\n

Agentic Coding Performance<\/strong><\/h2>\n

On SWE-Bench Pro<\/strong>, Step 3.7 Flash scores 56.26%<\/strong>, up from Step 3.5 Flash\u2019s 51.3% \u2014 a gain of roughly 5 percentage points. On Terminal-Bench 2.1<\/strong>, it scores 59.55%<\/strong>, up from 53.37%.<\/p>\n

On SWE-MTLG<\/strong> (a multi-task long-generation coding benchmark), it scores 72.42%<\/strong>.<\/p>\n

Cross-harness consistency on StepFun\u2019s internal Step-SWE-Bench<\/strong>:<\/p>\n

\n\n\n\n\n\n\n\n\n\n\n
Scaffold<\/th>\nStep 3.7 Flash<\/th>\nStep 3.5 Flash<\/th>\n<\/tr>\n<\/thead>\n
Hermes Agent<\/td>\n67.5%<\/td>\n60.0%<\/td>\n<\/tr>\n
OpenClaw<\/td>\n67.0%<\/td>\n47.0%<\/td>\n<\/tr>\n
KiloCode<\/td>\n67.5%<\/td>\n59.0%<\/td>\n<\/tr>\n
RooCode<\/td>\n64.5%<\/td>\n43.0%<\/td>\n<\/tr>\n
Claude Code<\/td>\n71.5%<\/td>\n73.0%<\/td>\n<\/tr>\n
OpenCode<\/td>\n64.5%<\/td>\n57.0%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

Step 3.5 Flash ranged from 43% to 73% across harnesses. Step 3.7 Flash ranges from 64.5% to 71.5%. In production, coding agents often run inside heterogeneous scaffolds \u2014 each with its own prompting conventions and tool schemas. Narrower per-harness variance means more predictable behavior across different setups.<\/p>\n

Advisor Mode<\/strong><\/h3>\n

Step 3.7 Flash supports Advisor Mode<\/strong>, StepFun\u2019s implementation of the advisor strategy described by Anthropic. The model runs the agentic loop end-to-end \u2014 calling tools, reading results, iterating \u2014 and escalates to a larger advisor model only at specific inflection points, such as planning or recovering from repeated failures. Most of the run stays at executor cost.<\/p>\n

With Advisor Mode enabled on SWE-Bench Verified, StepFun reports Step 3.7 Flash reaches 97% of Claude Opus 4.6\u2019s coding performance at roughly one-ninth the per-task cost<\/strong> ($0.19 vs. $1.76 per task). These are StepFun\u2019s internal figures.<\/p>\n

Multimodal Capabilities<\/strong><\/h2>\n

Step 3.7 Flash supports two visual tool pathways:<\/strong><\/p>\n

Visual Search Tool<\/strong> \u2014 For recognition tasks where the model\u2019s parametric knowledge is insufficient (long-tail entities, recently emerged concepts), it invokes a visual search tool to retrieve and verify. On SimpleVQA (with Search)<\/strong>, it scores 79.16%<\/strong>, comparable to GPT 5.5 (79.11%) and above Kimi K2.6 (78.24%) and GLM 5V Turbo (78.20%).<\/p>\n

Python Tool<\/strong> \u2014 For fine-grained visual tasks (high-resolution images, visual probing, bounding-box analysis), it uses a code interface to crop, zoom, and draw pixels or bounding boxes. On V<\/em> (a self-tested score with Python), it scores 95.29%<\/strong>. On HR-Bench 4K<\/strong> and HR-Bench 8K<\/strong>, it scores 89.13%<\/strong> and 86.34%<\/strong> respectively.<\/p>\n

StepFun notes an observed behavior during testing: the model combined visual tools with non-visual tools without being explicitly trained to do so. For example, after generating frontend code, it used the GUI to render and inspect the result before iterating. StepFun describes this as emergent compositional tool use.<\/p>\n

On Android Daily<\/strong> (long-horizon phone UI task completion), Step 3.7 Flash scores 61.87%<\/strong>, ahead of Kimi K2.6 (53.36%) and GLM 5V Turbo (51.68%). Gemini 3 Flash (63.21%) leads this benchmark.<\/p>\n

Search and Research Benchmarks<\/strong><\/h2>\n

StepFun focused this model\u2019s search design on planning, evidence filtering, and synthesis \u2014 integrating search as part of the reasoning loop rather than a separate add-on.<\/p>\n

\n\n\n\n\n\n\n\n\n
Benchmark<\/th>\nStep 3.7 Flash<\/th>\nNotable comparison<\/th>\n<\/tr>\n<\/thead>\n
HLE with Tools (acc)<\/td>\n47.20%<\/td>\nDeepSeek V4 Flash: 45.10%<\/td>\n<\/tr>\n
BrowseComp (acc)<\/td>\n75.82%<\/td>\nClaude Opus 4.7: 79.30%<\/td>\n<\/tr>\n
DeepSearchQA (F1)<\/td>\n92.82%<\/td>\nKimi K2.6: 92.50%<\/td>\n<\/tr>\n
ResearchRubrics (score)<\/td>\n71.68%<\/td>\nGPT 5.5: 61.50%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

Note: The HLE with Tools score of 47.20% compares to Step 3.5 Flash\u2019s text-only score of 35.68%. Step 3.5 Flash did not support tool-augmented evaluation on HLE.<\/p>\n

General Agent Benchmarks<\/strong><\/h2>\n
\n\n\n\n\n\n\n\n\n
Benchmark<\/th>\nStep 3.7 Flash<\/th>\nDescription<\/th>\n<\/tr>\n<\/thead>\n
Toolathlon<\/td>\n49.51%<\/td>\nMulti-tool coordination<\/td>\n<\/tr>\n
ClawEval-1.1<\/td>\n67.07%<\/td>\nDaily autonomous task execution in realistic environments<\/td>\n<\/tr>\n
GDPval (44 occupations)<\/td>\n45.8%<\/td>\nGeneral professional task execution<\/td>\n<\/tr>\n
Tau2-bench Telecom<\/td>\n>98%<\/td>\nAcross different reasoning difficulty tiers<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

On ClawEval-1.1, Step 3.7 Flash (67.07%) leads DeepSeek V4 Flash (57.80%) and DeepSeek V4 Pro (59.80%) among the compared models.<\/p>\n

Long-Context Performance<\/strong><\/h2>\n

On AA-LCR<\/strong> (a long-context retrieval benchmark, avg@16\/acc), Step 3.7 Flash scores 63.94%<\/strong>. This is comparable to DeepSeek V4 Flash (63.70%) and DeepSeek V4 Pro (66.30%).<\/p>\n

Pricing<\/strong><\/h2>\n
\n\n\n\n\n\n\n\n
Token Type<\/th>\nPrice<\/th>\n<\/tr>\n<\/thead>\n
Input (cache miss)<\/td>\n$0.20 \/ M tokens<\/td>\n<\/tr>\n
Input (cache hit)<\/td>\n$0.04 \/ M tokens<\/td>\n<\/tr>\n
Output<\/td>\n$1.15 \/ M tokens<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

Marktechpost\u2019s Visual Explainer<\/strong><\/h2>\n
\n

<\/p>\n

\n
Model Release<\/div>\n
Step 3.7 Flash \u2014 A 198B MoE Vision-Language Model<\/div>\n
StepFun \u00b7 Released May 29, 2026 \u00b7 Apache 2.0<\/div>\n<\/div>\n

<\/p>\n

\n
<\/div>\n<\/div>\n

<\/p>\n

\n

<\/p>\n

\n
Slide 1 of 8 \u2014 Overview<\/div>\n

What Is Step 3.7 Flash?<\/h3>\n

Step 3.7 Flash is a sparse Mixture-of-Experts (MoE)<\/strong> vision-language model from StepFun. It combines a 196B-parameter language backbone with a 1.8B-parameter Vision Transformer (ViT) encoder for native image understanding.<\/p>\n

In a MoE model, only a subset of \u201cexpert\u201d sub-networks activates per token \u2014 not the full network. This keeps inference compute close to an 11B dense model while maintaining 198B total parameters.<\/p>\n

\n
\n
Total Params<\/div>\n
198B<\/div>\n<\/div>\n
\n
Active \/ Token<\/div>\n
~11B<\/div>\n<\/div>\n
\n
Context Window<\/div>\n
256k tokens<\/div>\n<\/div>\n
\n
Throughput<\/div>\n
400 tok\/sec<\/div>\n<\/div>\n
\n
Reasoning Levels<\/div>\n
Low \/ Med \/ High<\/div>\n<\/div>\n
\n
License<\/div>\n
Apache 2.0<\/div>\n<\/div><\/div>\n<\/div>\n

<\/p>\n

\n
Slide 2 of 8 \u2014 Architecture<\/div>\n

Architecture Notes<\/h3>\n

The 1.8B ViT encoder runs as a separate module<\/strong> and injects image representations into the language backbone\u2019s context. Step 3.5 Flash was text-only; native multimodal support is new in 3.7.<\/p>\n

<\/div>\n

Three selectable reasoning depths<\/strong> let developers balance speed and cost:<\/p>\n