ModelScale open-sources MiniCPM5-2B with weights, data, RL framework — 2B beats 4B rivals
AGI Hunt · wechat · 2026-09-08
ModelScale (OpenBMB) has open-sourced MiniCPM5-2B, releasing not just weights but training data, RL framework, and recipes — a first for a major LLM vendor.
Benchmarks (ArtificialAnalysis)
- Intelligence Index of 23, topping Gemma 4 12B and Qwen3.5 9B (both 22) with 4-6x fewer parameters
- Agentic Index of 20 vs 9 for runner-up Granite 4 8B; GDPval-AAv2 Elo of 891 (human baseline 1000) vs 702 for second place
- Second-lowest output tokens per task (21k), so gains aren't from thinking longer; rated most honest in class, preferring abstention over guessing
What's open-sourced
- Four datasets: UltraData-Code, UltraData-SFT-Agent-2609 (1M+ agent trajectories), UltraData-RL-2609 (with a three-stage cleaning pipeline), and UltraX
- RL framework Meshy (no central controller, no Ray dependency; sync/async/fully-async modes) and JustRLII algorithm (GRPO with a Critic for token-level credit assignment)
Hands-on tests
- 50 synthetic resumes, 16-way concurrent extraction: 23s per round, 150/150 valid JSON, 97.1% field accuracy
- Agent task with a deliberately wrong filename: 5/5 runs self-recovered via listdir and completed filing in 12s
- Accurate recall at 125k tokens on a 33k-word contract QA test
The model is adapted to Intel, Rockchip, and Arm chips; the MiniCPM series has surpassed 50M downloads.
Related event: OpenBMB Open-Sources MiniCPM5-2B, Topping Sub-4B Open Models(14 posts)→
More from coding & agent
- BELL: Open-Source Stack Combines Browserbase, Eve, iMessage and Stripe Link for Personal Agents — jeff_weinstein · 2026-10-27
- Legora runs 2,000 parallel agent evals on Modal Sandboxes for legal AI — blelbach · 2026-10-27
- Two engineers rebuilt a FedEx supplier's billing core in 3.5 months vs an estimated 7-8 — alex_verem · 2026-10-27
- GPT-6.1 Terra scores 48 on AA Coding Agent Index at one-third the cost of GPT-5.6 Terra — ArtificialAnlys · 2026-10-27
- Shift-1 router cuts coding agent spend up to 55% while keeping SWE-bench accuracy — ypatil125 · 2026-10-27
- Non-coder vibe-codes a Godot JRPG with Claude Code in two weeks, music and art synced — ADHDavid · 2026-10-27