OpenBMB's MiniCPM-V 4.7 Went Up With No Model Card on Oct. 6, and Now Reads Apache-2.0 With No Benchmark Numbers
A.I. / news
OpenBMB's MiniCPM-V 4.7 Went Up With No Model Card on Oct. 6, and Now Reads Apache-2.0 With No Benchmark Numbers
The 35-billion-parameter mixture-of-experts vision model is 27 times larger than its predecessor, and its card claims a win over Qwen3.6 35B-A3B without a score.
OpenBMB's MiniCPM-V 4.7 35B-A3B now carries an Apache-2.0 licence on its Hugging Face model card, eight days after the repository appeared with no card at all. The card lists no benchmark scores in text, no context window and no release date.
The repository was created at 18:36 UTC on Oct. 6, according to OrcaRouter, a model-routing company that tracked the upload. Its last commit that evening was at 18:48 UTC. At that point, OrcaRouter reported, there was no README, no licence tag, no technical report and no mention in OpenBMB's GitHub README, which still named MiniCPM-V 4.6 as the latest model.
What the card says
The card calls it "the first MoE model in the MiniCPM-V series." MoE means mixture of experts, a design that routes each token to a subset of the model's parameters instead of all of them. It handles single images, multiple images and video, and outputs text.
It does not state the active parameter count. The "A3B" in the name points to about 3 billion active per token, but that is an inference from the name. OrcaRouter read the configuration file and found 256 experts with eight selected per token.
The card credits Qwen3.5 and Qwen3.6 for the language backbone and SigLIP2 for the vision module. It names no training data. It requires Transformers 5.18.0 or later and recommends Flash Attention 2 for multi-image and video work.
The claims that are OpenBMB's alone
The card says that in Instruct mode the model scores higher overall than Qwen3.6 35B-A3B, in less inference time. The benchmark charts are images, so no figure is in the page text. The comparison is vendor-supplied, and the card names no independent evaluator.
It also says the model runs at 13 or more tokens a second in about 3 GB of memory on an iPhone 17 Pro Max. That figure sits oddly beside the weights. OrcaRouter measured the BF16 checkpoint at 70,425,751,648 bytes across 16 shards, about 70.4 GB, so the phone number presumably refers to one of the quantized variants the card offers: FP8, int4, AWQ, GPTQ and GGUF. The card does not say which.
- MiniCPM-V 4.6 (May 11, 2026)1.3 B
- MiniCPM-V 4.7 35B-A3B35.2 B
Source: OrcaRouter, accessed 2026-10-14
Launch day against now
| Item | Oct. 6 launch | Card as fetched Oct. 14 |
|---|---|---|
| Licence | None declared | Apache-2.0 |
| Context window | 256K (262,144 positions, from config) | Not stated |
| Quantized builds | None | FP8, int4, AWQ, GPTQ, GGUF listed |
| Benchmarks | None | Charts as images, no text figures |
The card does not say whether the weights are gated, and it shows no gating notice. OrcaRouter, which sells a hosted routing service, framed the upload as staged for a launch. That is its speculation, and OpenBMB has not said so.
For scale, MiniCPM-V 4.6 shipped on May 11, 2026 at 1.3 billion parameters under Apache-2.0, built on a Qwen3.5-0.8B language model. The new model is roughly 27 times larger. It shares the Apache-2.0 licence of another open release the site has covered, Qwen3.8-27B, and lands in the same week as Gemini 4 Argon's pricing in the closed-model market.
OpenBMB has not published a technical report. The card mentions a separate draft model called DSpark for speculative decoding, and neither the card nor OrcaRouter gives a date for it.
Sources
More in A.I.
- 01Torvalds Posts a Kernel Patch Series Written With Claude Code, and Sources Disagree on Its Size and SpeedupLWN counts ten patches and an order-of-magnitude gain, while It's FOSS counts eight patches and a threefold gain on one laptop.
- 02Parsimony Ranks Coding Agents by How Much Code Their SWE-bench Fixes Add, and Its Own Page Warns Smaller Is Not BetterThe MIT-licensed benchmark counts coding units in published patches without rerunning a single test, and reports that code volume tracks solve rate.
- 03Design Creativity Bench Scores GPT-6 Astra 65.5 Against Humans' 92.3, and the Paper Never Names Its JudgeA post using Simon Willison's pelican SVGs to argue that models converge on one image cites a leaderboard whose method the arXiv page does not describe.
- 04Qwen3.8-27B Reaches 7.1 Million Downloads Under Apache 2.0, on Benchmarks Only Alibaba Has RunThe August 14 model beats Opus 4.6 Max on agentic coding by Alibaba's scores and trails it on reasoning. The card gives no training data and no GPU requirement.