# China's OpenBMB Releases MiniCPM5-2B, Beating Every Open Model Under 4B > Source: > Published: 2026-09-08 09:52:49+00:00 *OpenBMB's MiniCPM5-2B is small enough for local devices and strong enough to lead open-weight models below 4 billion parameters on Artificial Analysis's latest index. The harder part for rivals is that OpenBMB didn't stop at weights.* OpenBMB released MiniCPM5-2B on September 7. Skip its own model-card table - the cleanest number comes from Artificial Analysis. Artificial Analysis reported that the 2.6-billion-parameter dense reasoning model scored 15 on the Artificial Analysis Intelligence Index v4.2, the highest score it has measured for any open-weight model under 4 billion total parameters. Granite 4.2 3B was next at 11. That's the gap you should pay attention to. OpenBMB is a China-based open-source AI group jointly founded in 2022 by Tsinghua University's NLP Lab and ModelBest, according to Artificial Analysis. The new model supports text input and output, carries a 131,072-token context window, and ships under Apache 2.0 on Hugging Face. It's not a frontier model. Don't confuse the two. It is a small model that has made the old trade-off between size and usefulness look less fixed than it did a week ago. The footprint is the reason this release matters to builders rather than benchmark watchers. The official GGUF repository lists MiniCPM5-2B at 2,516,756,480 total parameters, with 1,981,982,720 outside the embedding layer. Its Q4_K_M build is 1.56 gigabytes, the Q8_0 build is 2.68 gigabytes, and the F16 file is 5.04 gigabytes. Those are laptop and phone-class numbers, not cloud-cluster numbers. You don't need a data center. You need a device that can keep a few gigabytes available and a runtime such as llama.cpp, Ollama, LM Studio, MLX on Apple Silicon, vLLM, or SGLang. For a founder trying to put AI into a field device, a car interface, a factory-floor terminal, or a phone app that can't assume a stable connection, that difference shows up as latency, battery use, and product design. Cloud APIs are still easier. Local models are becoming harder to dismiss. [Giving Your AI Agent an Email Address Is Now a Real Product Category](https://startupfortune.com/giving-your-ai-agent-an-email-address-is-now-a-real-product-category/) AI agents can send email easily, but receiving replies and holding a thread together is still the hard part. This piece compares seven email services and APIs, from AgentMail's newly funded inbox-first approach to Postmark, Resend, Mailgun, SendGrid, Amazon SES, and Mailtrap, on what actually works for autonomous agents. - [AI agents with email inbox functionality capabilities](https://startupfortune.com/giving-your-ai-agent-an-email-address-is-now-a-real-product-category/) - [email API platforms for autonomous agent communication workflows](https://startupfortune.com/giving-your-ai-agent-an-email-address-is-now-a-real-product-category/) ## The benchmark story needs a careful read The original temptation is to quote OpenBMB's own 53.9 average and call it an Artificial Analysis score. That's wrong. OpenBMB's Hugging Face model card reports a 53.9 average across its comparison table, ahead of Qwen3.5-4B at 51.1, but the card also says only the dagger-marked rows come from the official Artificial Analysis release. The Artificial Analysis index score is 15. Both numbers are useful. They aren't the same thing. That distinction matters because small-model headlines can get sloppy very quickly. OpenBMB's internal table shows attention-grabbing results including 69.1 on LiveCodeBench v6 and 97.1 on tau-squared-Bench Telecom. Artificial Analysis's own article is more measured: it says MiniCPM5-2B's agentic edge shows up in GDPval-AA v2, where it reached an Elo of 831, and in tau-cubed Banking, where it was joint-first with Ling 3.0 Tiny at 21%. Still, the outside result is strong enough. Artificial Analysis said MiniCPM5-2B used 19,000 output tokens per Intelligence Index task, tied for the lowest in its comparison set with Granite 4.2 3B. Ling 3.0 Tiny used 56,000 for one more index point. Token use is not trivia when the model is meant to run close to the user. It decides whether a local assistant feels quick or feels like a demo. ## The data release is the sharper move Here's the part closed labs should worry about. OpenBMB released more than model weights. Its model card says the training corpus includes Ultra-FineWeb, Ultra-FineWeb-L3, UltraX, UltraData-Code, and UltraData-Math for earlier training stages, plus UltraData-SFT-2605, UltraData-SFT-Agent-2609, and UltraData-RL-2609 for post-training. The UltraData-RL-2609 page on Hugging Face says the reinforcement-learning set contains about 86,000 samples across math, knowledge, long-context, and code tasks. That changes the value of the release. A startup can download weights. A rival lab can study the recipe. A researcher can inspect the data path instead of guessing what happened between pretraining and the final chat model. Apache 2.0 also keeps the commercial route open, so this isn't only an academic artifact sitting on Hugging Face to gather stars. Alibaba's Qwen line and DeepSeek have already shown how uncomfortable open-weight Chinese labs can make the model market for Western companies that keep weights and training details locked down. MiniCPM5-2B pushes that argument down to device scale. Frankly, a closed lab can match a leaderboard number faster than it can match the trust created by releasing the machinery behind the number. For you, the practical question is simple: does your product really need a larger hosted model for every task, or does it need a small local model that can call tools, handle long context, and stay cheap to run? MiniCPM5-2B doesn't answer that for every use case. It does make the lazy answer weaker. [Z.ai's New GLM-5.3-Flash Model Runs 3.3 Times Faster on a Single Workstation](https://startupfortune.com/zais-new-glm-53-flash-model-runs-33-times-faster-on-a-single-workstation/) Z.ai open-sourced GLM-5.3-Flash, a 320-billion-parameter MoE model that combines sparse and linear attention for a 3.3x inference speedup. It ships MIT-licensed with a 1M-token context window and can run locally on a single workstation with 102GB to 128GB of RAM, undercutting closed US labs on cost. - [GLM-5.3-Flash model speed improvements for workstations](https://startupfortune.com/zais-new-glm-53-flash-model-runs-33-times-faster-on-a-single-workstation/) - [mixture of experts model runs faster than predecessor](https://startupfortune.com/zais-new-glm-53-flash-model-runs-33-times-faster-on-a-single-workstation/) **Also read:** [AI Startups Are Counting Signed Contracts as Real ARR, VCs Admit](https://startupfortune.com/ai-startups-are-counting-signed-contracts-as-real-arr-vcs-admit/) • [SAP's Own Board Says It Needs an AI Breakthrough Within Months](https://startupfortune.com/saps-own-board-says-it-needs-an-ai-breakthrough-within-months/) • [Anthropic Walked Away From a $6 Billion Deal to Buy AI Startup Decart](https://startupfortune.com/anthropic-walked-away-from-a-6-billion-deal-to-buy-ai-startup-decart/) ## Join the discussion [Open in the community →](/community/) Almost there. Sign in and your reply posts straight away.