OpenBMB Ships MiniCPM5-2B With Its Full Training Stack

The 2B edge model averages 53.9 across 34 benchmarks; OpenBMB opens the data, recipes, and RL stack.

OpenBMB open-sourced MiniCPM5-2B, a 2B-parameter model built for the edge, with weights on Hugging Face, GitHub, and Modelscope. The lab reports it ranks first among open-source models under 4B parameters on the Artificial Analysis Intelligence Index with a score of 23, plus 20 on the Agentic Index. Across 34 benchmarks covering coding, math, long context, tool use, and agentic tasks it averages 53.9, ahead of the second-ranked 2B model at 33.2.

The release goes past the weights. OpenBMB is publishing the training data, recipes, and RL stack: UltraData-Code at about 550B tokens, UltraData-SFT-Agent-2609 at about 500K samples, UltraData-RL-2609 with 80K+ samples, and UltraX at about 100B tokens across 114M samples. MiniCPM5-2B ships day-0 adapted for Intel Core Ultra with OpenVINO, Armv9 with SME2, and Rockchip's RK3588 with RK1828. On Arm SME2 mobile silicon the lab measures about 1.7× higher prefill and 1.2× higher decode.

The ecosystem met it on day one. @sgl_project confirmed MiniCPM5-2B runs on RTX PRO 6000, RTX 5090, and DGX Spark with SGLang. @aijoey congratulated the team; @naidx0 noted an 8GB GPU can now run it. OpenBMB's framing for the size class: “2B parameters doesn't mean 2B-level capability.” The posted token-budget table gives MiniCPM5-2B 21K tokens of thinking and answer against 19K for Granite 4.2 3B.

From @OpenBMB

Agents & Coding

Hermes Review: A Hackable Personal Agent With Gaps

Nous Research's agent is #1 on OpenRouter by a wide margin; @thomasgauvin finds polish and DIY rough edges.

@thomasgauvin reviewed Hermes, the Nous Research personal agent that also codes: polished desktop app and TUI, scheduled jobs, computer use and Chrome extensions out of the box, and the Grok bot rebuilt inside it. It is “#1 most popular app on openrouter by a large margin.” He flags no mobile app or remote mode and pre-baked skills of unknown provenance. @Teknium took the docs complaint himself: the default Docusaurus theme is his work.

From @thomasgauvin

An Anthropic Prompt Lifts Writing on Fable 5.1 and Sol

@omarsar0 says it improved Fable 5.1 output and, surprisingly, GPT-5.6 Sol output too.

@omarsar0 reports that a prompt Anthropic shared “significantly improved writing with Fable 5.1” and, more surprisingly, also improved GPT-5.6 Sol. He keeps it as a standing rule wherever he uses AI for editing or writing and built an interactive tutorial on the DAIR Academy so readers can test the difference themselves. His advice: try it once and watch the output closely.

From @omarsar0

Models & Hardware

Kearuga Recipe Holds 262K Context on One Spark, Near BF16

Tiered precision halves first-token KL divergence to 0.0165 against a served BF16 reference.

@yume_arasaki tested @0xWhiteMage's Qwen3.8-27B Kearuga build on a single DGX Spark: 24.85 GB of tiered precision, GPTQ-4o6 and NVFP4 where the model tolerates it, FP8 and BF16 where it does not, with logit tails and draft taps kept. First-token KL against a served BF16 reference fell from 0.0334 to 0.0165, with 40/40 top-1 agreement on a 40-prompt verification suite. His receipts: 57.7 tok/s on code, 39.4 agentic, and a 3/3 needle pass at 254,631 prompt tokens.

From @yume_arasaki

Two Agents Log 443 Tool Rounds in 24 Hours on TP-2 Qwen

Qwen 3.8 Flash Next NVFP4 across two DGX Sparks now carries Jinx and Luna in constant continuity.

@jmacftw moved his agents Jinx and Luna to Qwen 3.8 Flash Next NVFP4 in TP-2 across two DGX Sparks. Earlier tests kept the pair on DeepSeek 4 Flash: Qwen then had correctness and stability issues, and GLM 5.3 Flash could not carry the concurrent load. In the latest 24-hour window the two completed 443 tool-using rounds, 116 with verified changes. Newer weights lifted median short prose generation from 30.09 to 41.97 tokens/s, a 39.46% gain.

From @jmacftw

Aeon Forge Ships Official NVFP4 Quant of Qwen3.8-27B

Mixed precision per layer, calibrated for Blackwell hardware, up on Hugging Face.

@SpaceTimeViking released the long-awaited official calibrated NVFP4 quant of Qwen3.8-27B AEON ULTIMATE UNCENSORED, calling it “a careful use of mixed precision on each layer where it counts most.” The model card is live on Hugging Face as AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-NVFP4-MIXED, and @0xWhiteMage praised the quant's build quality. @Stonewrot says he may try it on a dual 5060ti rig.

From @SpaceTimeViking

Sources

Sunday, September 6, 2026 Back to the Archive Tuesday, September 8, 2026