MiniCPM5-2B beats every open model under 4B — and runs on your phone
The 2.52B Apache-2.0 model averages 53.9 across 34 benchmarks with day-one GGUF and MLX builds. Plus: AllSpark's open Iris search agents hit the frontier.
Copy markdown
2.52B params, and it tops the sub-4B leaderboard
OpenBMB's MiniCPM5-2B is a dense 2.52B model under Apache 2.0 with a 131K-token context. It averages 53.9 across 34 benchmarks — ahead of every open model in the 4B-and-under class, whose best manages 51.1 — and leads on coding, math, long-context, and tool use.
Day-one GGUF and MLX — it fits on-device
The launch ships GGUF, 4-bit GPTQ, and Apple MLX builds at once, plus support across vLLM, SGLang, llama.cpp, Ollama, LM Studio, and Docker Model Runner. At ~2.5B it quantizes small enough for a laptop or phone — a genuinely capable local model to wire into an offline agent this weekend. A DSpark draft model ships for speculative decoding.
Iris: open-weight search agents at the frontier
AllSpark Research open-sourced Iris-pro (397B/17B active) and Iris-mini (35B/3B), post-trained deep-research agents built on Qwen3.5/3.6. Iris-pro scores 88.6 on BrowseComp — the strongest open result to date — and Iris-mini hits 82.2, with weights and a 256K context on Hugging Face. You can build a browsing/research agent on open weights instead of a closed API.
Elsewhere: Meta's Muse agent opens a closed alpha
Meta began closed-alpha testing of Muse, its consumer agentic assistant, gated behind a waitlist and invite codes. It is pitched to take everyday actions on your behalf — the thing to watch is the permission model it normalizes, not access you can get today.