brain/
← all entities
entitygenericartificial-intelligence

MiniCPM5-2B

Notes

Vintage: 2026-09. Primary evidence is one Grok Bot midday X + fetch pass in 2026-09-07-openbmb-minicpm5-2b-open-release-aa-index-15 (method: grok-bot; x_video: false). Grain is the fetched Hugging Face card, the fetched Artificial Analysis model page, and official @OpenBMB / @ArtificialAnlys / @sgl_project X permalinks. ModelScope / openbmb.cn were linked and not body-fetched. No video in this set.

MiniCPM5-2B

One-line summary: openbmb's 7 Sep 2026 Apache-2.0 dense on-device LM; Hugging Face lists 2,516,756,480 parameters and 131,072 context; Artificial Analysis Intelligence Index v4.2 = 15 (under-4B open lead). Launch-post Index 23 is a superseded claim — keep it on the record.

What it is

A dense LlamaForCausalLM open-sourced by OpenBMB on 2026-09-07. Treat HF + AA pages as fetched issuer/independent cards; treat X as primary/discourse. This is not a frontier-parity crossing and not a rewrite of gemini-3-8-flash.

Why it matters to this thread

Open-weight / on-device small-model drops are in-scope (chinese-open-weight-frontier-parity, open-vs-closed-source-model-economics, edge-inference-shift). This is the first dated MiniCPM5-2B page. It does not rewrite the June GLM 5.2 / July Kimi K3 podcast evidence. It does not invent a "small-model" concept.

Key facts (from 2026-09-07-openbmb-minicpm5-2b-open-release-aa-index-15)

Issuer card and launch

  • From 2026-09-07-openbmb-minicpm5-2b-open-release-aa-index-15 (@OpenBMB, 2026-09-07): OpenBMB announced MiniCPM5-2B as open source, pointing to Hugging Face, GitHub, ModelScope, and openbmb.cn.
  • From the same source (fetched Hugging Face card): dense LlamaForCausalLM; 2,516,756,480 parameters; 131,072 context; Apache-2.0 weights; formats include BF16, GGUF, MLX, GPTQ, and a DSpark draft model.
  • From the same source (HF card + @OpenBMB): 53.9 average across a 34-benchmark internal set (2B-class SOTA in that table; second-listed 2B peer 33.2; listed 4B-class peers headed by Qwen3.5-4B at 51.1). 53.9 is 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.
  • From the same source (@OpenBMB): 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). Named artifacts, not new product pages.
  • From the same source (@OpenBMB): day-0 adaptation claimed 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 Index 15 vs launch Index 23

  • From the same source (@ArtificialAnlys, 2026-09-07): Intelligence Index v4.2 = 15; 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.
  • From the same source (fetched AA model page): Index 15; release date September 7, 2026; 2.6B parameters; 131k context; Apache 2.0; reasoning variant; knowledge cutoff Dec 31, 2025.
  • From the same source (launch post): the model "ranks #1 … on the @ArtificialAnlys Intelligence Index, with a score of 23" and "20 on the Agentic Index."
  • From the same source (@OpenBMB): later the same day OpenBMB quote-thanked AA and stated Index v4.2 = 15 — aligning with AA, not with the launch 23.

Prefer 15 for AA Intelligence Index citations. Keep launch 23 (and Agentic Index 20) on the record. Do not invent a reconciliation (do not assume 23 was a typo for Agentic).

Ecosystem day-0 (claims, not re-measured)

  • From the same source (@sgl_project): 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.
  • From the same source: OpenBMB thanked SGLang (2097007177315860627) and vLLM (2097008666071429495) for day-0 support. No new SGLang / vLLM product pages.

What this source does not establish

  • Not a frontier-parity crossing. Under-4B open lead at Index 15 is not GLM 5.2 / Kimi K3. See chinese-open-weight-frontier-parity.
  • 53.9 is OpenBMB's internal 34-bench table, not the AA Intelligence Index.
  • Launch Index 23 is not the independent Index figure. Prefer 15.
  • Parameter rounding is not a second model. HF 2,516,756,480; AA 2.6B; OpenBMB marketing "2B-parameter" / "2.6B" in the AA thank-you. Cite the exact figure with its source.
  • ModelScope / openbmb.cn not body-fetched. HF + AA + X were primary.
  • SME2 ~1.7× / ~1.2× and SGLang >250 tok/s/user are issuer / ecosystem claims, not independent benches.
  • Did not re-file gemini-3-8-flash. Clip skipped Gemini 3.8 discourse (already ingested 2026-09-03).
  • No video. x_video: false. Photos only. No transcripts invented.
  • Quiet labs after 14:00Z (@OpenAI / @AnthropicAI / @GoogleDeepMind / @GeminiApp / @AIatMeta / @karpathy / @sama / @demishassabis) is a negative finding, not grain.
  • PsAIch / "When AI Takes the Couch" was discourse-only (no paper fetch).
  • No new products. UltraData / Meshy / JustRL II / DSpark / SGLang / vLLM stay named artifacts on this page.

Contradictions / tensions

  • 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." Filed both; prefer 15 for Index grain. Not reconciled (user adjudication: keep both; do not assume 23 = Agentic).
  • Parameter rounding. Same model family; three figures. Not flattened.
  • 53.9 vs 15. Different instruments. Do not collapse.

Open questions

  • Whether Ling 3.0 Tiny remains the under-4B Index leader at 16 after further AA refreshes.
  • What ModelScope / openbmb.cn pages add beyond the fetched HF card.
  • What "Agentic Index 20" measures, and whether it is independent of the Intelligence Index.
  • Whether Index v4.2 = 15 moves under Index v4.3. From 2026-09-07-artificial-analysis-intelligence-index-v4-3: not re-checked this pass. Do not restate 15 as a v4.3 reading.

Sources

Related

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
DeepSeek-V4.1-FlashOpenBMB