Step-3.7-Flash
PulseAugur coverage of Step-3.7-Flash — every cluster mentioning Step-3.7-Flash across labs, papers, and developer communities, ranked by signal.
- 2026-07-20 product_launch A new AI model, Step 3.7 Flash, has been released, claiming superior performance and cost-efficiency compared to existing models. source
- 2026-06-18 product_launch StepFun launched Step 3.7 Flash, an upgraded LLM with enhanced vision capabilities and an automatic task escalation feature. source
- 2026-06-05 product_launch Step 3.7 Flash was released and achieved top rankings on the AA benchmark. source
- 2026-06-05 product_launch Step (Jieyue) launched its Step 3.7 Flash model, which achieved top rankings for speed, cost-efficiency, and end-to-end performance. source
- 2026-06-04 research_milestone StepFun's Step 3.7 Flash model achieved a leading output speed of 409 tokens/s on the Artificial Analysis platform. source
- 2026-06-04 product_launch StepFun released its new open-source model, Step 3.7 Flash, which achieved top performance in AI output speed benchmarks. source
- 2026-06-01 product_launch Fireworks AI has launched the Step 3.7 Flash model, emphasizing its inference efficiency. source
- 2026-05-30 product_launch StepFun launched its Step 3.7 Flash model, claiming near-parity with Claude Opus 4.6 in coding performance at a significantly lower cost. source
- 2026-05-29 product_launch StepFunai has released the Step-3.7-Flash model with day-zero support in vLLM. source
- 2026-05-29 product_launch StepFun released Step 3.7 Flash, a 198B parameter MoE vision-language model for agentic workflows. source
- 2026-05-29 product_launch JieYue released and open-sourced the Step 3.7 Flash model. source
- 2026-05-23 product_launch Stepfun AI released the Step 3.7 Flash, a new multimodal MoE model optimized for agentic workflows. source
6 day(s) with sentiment data
-
Step 3.7 Flash claims 23x more intelligence per dollar than Claude Fable 5
A new AI model, Step 3.7 Flash, claims to offer 23 times greater intelligence per dollar compared to Anthropic's Claude Fable 5. The model boasts a higher intelligence index (59.9 vs. 30.3) and a significantly lower ble…
-
Eider inference runtime launched for NVIDIA DGX Spark, bypassing existing libraries
A new inference runtime called Eider has been developed for NVIDIA DGX Spark and similar hardware, built from scratch in Rust and CUDA. Eider is designed to leverage the NVFP4 capabilities of SM121 GPUs and does not rel…
-
LM Arena scales back inclusion of new open-source LLMs
The LM Arena, a platform for comparing open-source large language models, appears to be reducing its inclusion of newly released open models. Users have noted that the platform is no longer displaying many recent open m…
-
Local LLM optimization: Step-3.7-Flash gains 2.4x speed, MTP breaks vision
A developer has optimized the Step-3.7-Flash (198B-A11B vision MoE) model for local hardware, achieving significant performance gains. By ensuring the model's largest quantization (IQ3_XXS) fits entirely within the 96GB…
-
Mimo 2.5 excels at large context tasks on consumer GPUs
The Mimo 2.5 large language model demonstrates impressive speed and performance with large context windows, particularly on dual RTX Pro 6000 GPUs. This is attributed to its efficient 5-to-1 local/global sliding-window …
-
Qwen 3.7 unlikely to be open-sourced amid team departures
Speculation suggests that Qwen will not release their Qwen 3.7 model as open source, following the departure of key personnel like Junyang Lin. This move would position Qwen as the last major Chinese AI lab to not have …
-
StepFun releases Step 3.7 Flash with vision and auto-escalation
StepFun has released Step 3.7 Flash, an upgraded version of its 3.5 Flash model, featuring a new vision encoder and an automatic "Advisor Mode" that escalates complex tasks to larger models. This update aims to improve …
-
Step-3.7-Flash on AMD/ROCm faces context corruption and requires thinking budget
A user running the Step-3.7-Flash model on AMD hardware with ROCm has identified two key issues. First, ROCm appears to corrupt context windows beyond approximately 94,000 tokens, causing the model to loop and fail to p…
-
Stepfun's Step 3.7 Flash matches Claude's coding at 1/9 cost
Chinese AI startup Stepfun has released its Step 3.7 Flash model, which reportedly achieves 97% of Claude's coding capabilities at one-ninth the cost. This new model offers a speed of 416 tokens per second, making it a …
-
Step 3.7 Flash leads benchmarks in speed, cost, and performance
StepFun's new model, Step 3.7 Flash, has achieved top rankings on the Artificial Analysis (AA) benchmark, excelling in speed, cost-efficiency, and end-to-end performance. The model demonstrates impressive output speeds …
-
StepFun's Step 3.7 Flash leads AI output speed benchmarks
StepFun's new open-source model, Step 3.7 Flash, has achieved the top position on Artificial Analysis's Output Speed leaderboard, processing 409 tokens per second. The model also leads in other key metrics like end-to-e…
-
AliExpress 618 Sale Sees Near 40% Brand GMV Penetration
AliExpress's summer 618 promotion, which began on June 1st, has seen a significant surge in brand Gross Merchandise Volume (GMV) penetration, reaching nearly 40% on its opening day. This event has facilitated a concentr…
-
Fireworks AI releases 196B MoE model optimized for inference
Fireworks AI has released Step 3.7 Flash, a 196-198 billion parameter Mixture-of-Experts (MoE) model. This model was specifically designed with inference efficiency in mind from its inception. The company highlights tha…
-
StepFun's Flash Model Nears Claude Opus Coding Performance at Fraction of Cost
StepFun has released its Step 3.7 Flash model, which reportedly achieves 97% of the coding performance of Anthropic's Claude Opus 4.6. This new model is significantly more cost-efficient, operating at one-ninth the cost…
-
Haiguang and Jieyue Xingchen optimize AI model for efficient deployment
Jieyue Xingchen has released its Step 3.7 Flash model, which has been fully adapted and optimized by the Haiguang team. This integration allows developers to efficiently run the model on Haiguang's platform, facilitatin…
-
StepFunai releases 198B sparse MoE vision-language model
StepFunai has released Step-3.7-Flash, a 198 billion parameter sparse Mixture-of-Experts model. This new vision-language model offers day-zero support within the vLLM inference engine. The integration with vLLM is highl…
-
RTX 6000 Blackwell GPUs tested with Step 3.7 Flash
A user on Reddit shared early data and configuration details for running Step 3.7 Flash on two Blackwell RTX 6000 graphics cards. The post includes initial readings on tokens per second for general inference and links t…
-
StepFun releases 198B MoE vision-language model for agents
StepFun has released Step 3.7 Flash, a 198 billion parameter Mixture-of-Experts vision-language model designed for coding agents and search workflows. This new model features native multimodal understanding, improved to…
-
Stepfun AI releases 198B parameter multimodal MoE model
Stepfun AI has released Step 3.7 Flash, a 198-billion parameter sparse Mixture-of-Experts (MoE) vision-language model. This model is optimized for agentic workflows, coding, and multimodal tasks, activating approximatel…