brain/
sourceartificial-intelligence

OpenBMB MiniCPM5-2B open release + AA Index 15 (2026-09-07)

Post-10am X pass: OpenBMB ships MiniCPM5-2B (dense ~2.6B, Apache 2.0, 131k); AA Intelligence Index v4.2 = 15 under-4B open lead; launch post's Index 23 later aligned to 15.

Source

OpenBMB MiniCPM5-2B open release + AA Index 15 (2026-09-07)

Generated by Grok Bot research on 2026-09-07. WebSearch + fetch ladder + native X. Treat as raw material — review before promoting into a project or thread.

Summary

On 2026-09-07 OpenBMB open-sourced MiniCPM5-2B, a dense on-device language model (~2.52B parameters on the Hugging Face card; Artificial Analysis lists 2.6B), 131k context, Apache 2.0, with accompanying UltraData datasets and an RL stack. Artificial Analysis scores it 15 on Intelligence Index v4.2 and calls it the top open-weights model under 4B total parameters; OpenBMB later thanked AA and restated that 15. The same day's launch post had claimed an Intelligence Index score of 23 (plus Agentic Index 20) — treat 15 as the independent/issuer-aligned Index figure and keep 23 as a superseded launch claim. Issuer-table average 53.9 across 34 internal benchmarks is a separate OpenBMB comparison, not the AA Index.

Findings

Model drop (issuer)

OpenBMB announced MiniCPM5-2B as open source on X (@OpenBMB, 2026-09-07), pointing to Hugging Face, GitHub, ModelScope, and openbmb.cn. The Hugging Face model card describes a dense LlamaForCausalLM with 2,516,756,480 parameters, 131,072 context, and Apache-2.0 weights; formats include BF16, GGUF, MLX, GPTQ, and a DSpark draft model. The card and thread claim a 53.9 average across a 34-benchmark internal set (2B-class SOTA in that table; second-listed 2B peer 33.2; also above listed 4B-class peers headed by Qwen3.5-4B at 51.1). A follow-up post (@OpenBMB) says the release opens training data and stack pieces (UltraData-Code ~550B tokens; UltraData-SFT-Agent-2609 ~500K samples; UltraData-RL-2609 80K+ samples; UltraX ~100B tokens / 114M samples; Meshy; JustRL II). Another follow-up (@OpenBMB) claims day-0 adaptation on Intel Core Ultra + OpenVINO, Armv9 + SME2 (~1.7× prefill / ~1.2× decode on SME2 mobile, issuer claim), and Rockchip RK3588 / RK1828, plus SGLang / vLLM / llama.cpp / Ollama / Transformers support.

Independent AA score vs launch Index claim

Artificial Analysis reported MiniCPM5-2B at 15 on Intelligence Index v4.2, highest among open-weights models under 4B total parameters; one point behind Ling 3.0 Tiny (16, ~3× parameters); GDPval-AA v2 Elo 831 leading the <4B set in that write-up; ~19k output tokens per Index task. The AA model page (artificialanalysis.ai/models/minicpm5-2b) repeats Index 15, release date September 7, 2026, 2.6B parameters, 131k context, Apache 2.0, reasoning variant, knowledge cutoff Dec 31, 2025. OpenBMB's launch post (same thread) said the model "ranks #1 … on the @ArtificialAnlys Intelligence Index, with a score of 23" and "20 on the Agentic Index." Later the same day OpenBMB quote-thanked AA and stated Index v4.2 = 15 (@OpenBMB) — aligning with AA, not with the launch 23. Prefer 15 for AA Intelligence Index citations; do not silently collapse launch 23 into the Index without noting the walk-back.

Ecosystem day-0

SGLang claimed day-0 serving recipes (RTX PRO 6000, 5090, DGX Spark, Hopper) and ">250 tok/s/user decode at bs=1 on coding tasks with DSpark enabled on a 5090" (ecosystem claim, not re-measured here). OpenBMB thanked SGLang (@OpenBMB) and vLLM (@OpenBMB) for day-0 support.

Out of scope this pass

Official @OpenAI / @AnthropicAI / @GoogleDeepMind / @GeminiApp / @AIatMeta / @karpathy / @sama / @demishassabis timelines had no original posts after 2026-09-07T14:00:00Z. Gemini 3.8 Flash discourse was skipped — already ingested (vault/threads/artificial-intelligence/sources/2026-09-03-gemini-3-8-flash-and-3-8-flash-cyber-sep-2-2026.md). PsAIch / "When AI Takes the Couch" cluster was discourse-only here (no paper fetch). No video posts in the MiniCPM set (x_video: false).

Contradictions and open questions

  • AA Intelligence Index 15 vs OpenBMB launch claim 23: AA post + AA page + OpenBMB's later thank-you agree on 15. Launch post still reads 23 on the Intelligence Index and 20 on an "Agentic Index." File both; prefer 15 for Index grain; do not invent a reconciliation (e.g. do not assume 23 was a typo for Agentic without issuer saying so).
  • Parameter rounding: HF card 2,516,756,480; AA 2.6B; OpenBMB marketing "2B-parameter" / "2.6B" in the AA thank-you. Same model family; cite the exact figure with its source.
  • 53.9 average is OpenBMB's internal 34-bench table on the HF card, not the AA Intelligence Index. Several † rows on that table are marked as coming from Artificial Analysis; do not treat the whole 53.9 as an AA composite.
  • Whether Ling 3.0 Tiny remains the under-4B Index leader at 16 after further AA refreshes is open.
  • ModelScope / openbmb.cn pages were linked but not fully fetched this pass (HF + AA + X were primary).

Provenance

Method: Grok Bot / WebSearch / fetch ladder / native X Generated: 2026-09-07 Rounds: 1 of 3 — early-exit: single bounded midday model drop; labs quiet after 10am ET URLs fetched: HF card + AA model page successful via WebFetch; X via user-X MCP (search_news pointer only, get_users_posts, get_posts_by_ids); ModelScope/openbmb.cn not body-fetched Fetch notes: News-cluster summaries used only as pointers to post IDs — not cited as grain. search_posts_all skipped (user-OAuth). Deduped against same-day AI-thread X sources 2026-09-07-x-overnight-openai-rsi-research-acceleration-data-jensen-agi and 2026-09-07-weekend-x-ai-meta-aira3-kaggle-gold-openai-wiki.

Web sources:

X sources:

Grokipedia:

  • not used
Referenced by
Concepts
Artificial Analysis Intelligence IndexChinese open-weight models approach frontier parityFrontier-class intelligence is escaping the datacenter to the edgeOpen vs closed source model economics
Entities
MiniCPM5-2BOpenBMB