--- title: "Nous Research Drops Step 3.7 Flash Free Inside the World's Fastest-Growing Agent" subtitle: "StepFun's 198B MoE vision-language model is now free on Nous Portal for 30 days, giving Hermes Agent users a powerful new backbone for coding, search, and multimodal workflows." url: "https://alphasignal.ai/news/nous-research-drops-step-3-7-flash-free-inside-the-world-s-fastest-growing-agent" author: "AlphaSignal Newsroom" author_url: "https://alphasignal.ai/editorial-team" company: "Nous Research" company_url: "https://alphasignal.ai/company/nousresearch" published: "2026-05-30T13:59:17.000Z" updated: "2026-06-16T20:48:06.919Z" type: "News" topics: - "Llms" - "Api" subtopics: - "Mixture Of Experts" - "Vision Language" upvotes: 1655 read_time: "6 min" source_url: "https://x.com/NousResearch/status/2060722721516872020" comments: 0 --- # Nous Research Drops Step 3.7 Flash Free Inside the World's Fastest-Growing Agent > StepFun's 198B MoE vision-language model is now free on Nous Portal for 30 days, giving Hermes Agent users a powerful new backbone for coding, search, and multimodal workflows. [Nous Research](https://portal.nousresearch.com/signup) just handed its Hermes Agent users a significant free upgrade: [Step 3.7 Flash](https://huggingface.co/stepfun-ai/Step-3.7-Flash), the newest model from Chinese frontier lab StepFun, is now available at no cost for 30 days through Nous Portal. For an agent platform that has been quietly becoming the most-used app on OpenRouter, this is a meaningful model drop , and it signals a growing cross-lab partnership between two of the most interesting teams in open AI right now. ## What just landed Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model. It pairs a 196B-parameter language backbone with a 1.8B-parameter Vision Transformer encoder for native image understanding, and activates approximately 11B parameters per token during inference. That last number is the key one. In MoE architectures, only a subset of "expert" sub-networks fires per forward pass , not the full network. This keeps inference compute closer to an 11B dense model while maintaining a 198B total parameter budget. Engineered for high-frequency production workloads, it delivers throughput of up to 400 tokens per second and supports a 256K context window. It also offers three selectable reasoning levels , low, medium, and high , so developers can easily balance speed, cost, and cognitive depth. ## Built for agents, not just chat This model was not designed to win chatbot leaderboards. It is designed to handle intensive tasks such as parsing massive financial reports in one pass, running multi-step search loops with cross-source verification, or operating concurrent coding agents in high-throughput pipelines. StepFun also baked in a novel "Advisor Mode" , the model runs the agentic loop end-to-end, calling tools, reading results, and iterating, then escalates to a larger advisor model only at specific inflection points like planning or recovering from repeated failures. Most of the run stays at executor cost. With Advisor Mode enabled on SWE-Bench Verified, StepFun reports Step 3.7 Flash reaches 97% of Claude Opus 4.6's coding performance at roughly one-ninth the per-task cost. The benchmark picture is compelling across the board: - 74.4% on SWE-bench Verified, 51.0% on Terminal-Bench 2.0, 88.2 on τ²-Bench - 56.3 on SWE-Bench PRO, placing second overall behind only Claude 4 Opus at 64.3. On ClawEval-1.1, it takes first place at 67.1, a full 7.3 points ahead of the next closest competitor. - 47.20% on HLE with Tools, up from 35.68% (text-only) for Step 3.5 Flash, outperforming Flash models from DeepSeek V4 and Gemini 3.5. - 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. The architecture also features hybrid attention that interleaves Sliding Window Attention (512-token window) and Global Attention at a 3:1 ratio, plus 3-way Multi-Token Prediction (MTP-3) for low-latency reasoning chains. On the pricing side, it runs at $0.20 per million input tokens and $1.15 per million output tokens, with a 256,000-token context window. ## Who is StepFun? If you haven't heard of StepFun, you're not alone , it's one of the most capable but least Western-press-covered Chinese frontier labs. Founded in April 2023 by Jiang Daxin, a former Global Vice President and Chief Scientist at Microsoft Software Technology Center Asia, the company focuses on building large language models and multimodal AI systems. StepFun is one of China's "Six Little Tigers" of AI, a group of six prominent AI startups that also includes Zhipu AI, Moonshot AI, MiniMax, Baichuan Intelligence, and 01.AI. In January 2026, StepFun announced a B+ round exceeding RMB 5 billion (USD $717 million), with state-backed investors like Shanghai State-owned Capital Investment and China Life Insurance, alongside returning shareholders Tencent. Reports in early 2026 referenced StepFun preparing a Hong Kong IPO targeting up to $500 million, with a possible pre-IPO valuation in the $6 billion to $10 billion range. This is a well-resourced lab making a deliberate push into the Western developer ecosystem through partnerships like this one with Nous. ## The Hermes Agent context The partnership matters more when you understand how fast Hermes Agent has grown. The open-source self-improving AI agent from Nous Research has accumulated over 180,000 GitHub stars in under four months since its February 25 launch, making it the fastest-growing open-source agent framework of 2026 by Dealroom's count. Hermes Agent now processes more tokens on OpenRouter than any other app in the world , 271 billion per day. On May 6, 2026, Nous Research announced it had reached number one on OpenRouter's global token rankings, surpassing OpenClaw's 186 billion daily. Hermes Agent is an open-source autonomous AI agent built by Nous Research. It's not a coding copilot tethered to an IDE or a chatbot wrapper around a single API. Hermes lives on your server, remembers what it learns, and gets more capable the longer it runs. Its self-improvement mechanism, called GEPA, is an ICLR 2026 Oral-accepted technique that makes agents with 20 or more self-generated skills 40 percent faster on repeated tasks. ## Nous Portal: the one-subscription pitch Nous Portal launched as a subscription product that bundles 300+ models and built-in tools , web search, image generation, TTS, browser automation , behind a single login. The value proposition is consolidation: without the gateway, hooking each of those tools up means a Firecrawl account, a FAL account, a Browser Use account, an OpenAI key, and a Modal account , five separate signups, five separate dashboards, five separate top-up flows. With the gateway, all of it routes through one subscription. The tool backends are worth knowing: web search uses Firecrawl, image generation uses FAL, text-to-speech uses OpenAI, and the cloud browser uses Browser Use. Portal tiers are Free, Plus, Super, and Ultra. Importantly, Step 3.7 Flash's 30-day free access is available even on the free tier , a low-friction way to test a frontier-class model inside a full agentic workflow. If you'd rather skip Portal entirely, that's also an option. Hermes supports 20+ LLM providers, including OpenRouter, direct Anthropic/OpenAI APIs, local Ollama, and vLLM. Bring your own API keys and skip Portal entirely. ## What this means in practice The real story here is about how the agentic model market is evolving. StepFun gets distribution inside the fastest-growing agent framework in the world. Nous gets a frontier-class multimodal model to offer its users for free, making the Portal subscription more compelling. And users get a 198B-parameter model with vision, 256K context, and tool-calling built in , at zero cost for a month. For inference and serving, developers can use vLLM, SGLang, Hugging Face Transformers, and llama.cpp. For model development and customization workflows, StepFun model support has also landed in the NVIDIA NeMo ecosystem, including AutoModel, Megatron Core, and Megatron Bridge. A model this capable, with Apache 2.0 licensing, that fits on a single 128GB device changes what is possible for on-premise AI deployments. The 30-day free window on Nous Portal is the easiest way to put it through its paces before committing to anything. ## Sources - [X post by @nousresearch](https://x.com/nousresearch/status/2060722721516872020) - [StepFun Step-3.7-Flash on Hugging Face](https://huggingface.co/stepfun-ai/Step-3.7-Flash) - [StepFun Releases Step 3.7 Flash – MarkTechPost](https://www.marktechpost.com/2026/05/29/stepfun-releases-step-3-7-flash-a-198b-moe-vision-language-model-for-coding-agents-and-search-workflows/) - [Step 3.7 Flash on NVIDIA Build](https://build.nvidia.com/stepfun-ai/step-3.7-flash/modelcard) - [Nous Portal – Hermes Agent Docs](https://hermes-agent.nousresearch.com/docs/integrations/nous-portal) - [Hermes Agent Desktop App – Medium](https://medium.com/@tentenco/hermes-agent-desktop-app-everything-you-need-to-know-about-nous-researchs-self-improving-ai-agent-3cb59bd31e5f)