Draft PR in llama.cpp, reported by a detection bot
Support for an unreleased model bearing the name of Chinese AI research group OpenBMB, "MiniCPM5-15B-A2.5B," was submitted on the afternoon of Nov. 16 Japan time as a draft PR (pull request, a proposed change) to the open-source inference software "llama.cpp." llama.cpp is open-source software for running AI models on local devices such as personal computers. The PR is titled "xcy/models: add MiniCPM5 MoE inference support," and as of early Nov. 17 it remained a draft and had not been merged. OpenBMB has made no announcement, and the information is an unconfirmed leak.
The X detection bot @aitrackerbot, which automatically tracks AI model releases, reported the PR on Nov. 16 (early Nov. 17 Japan time). The bot said OpenBMB is preparing a MiniCPM5 "MoE" (a design that runs only some of several smaller specialized networks) model, with 15 billion total parameters and 2.5 billion active. It described the PR's author as "A ModelBest engineer" and also wrote that the PR was tested against GGUFs (the file format used by llama.cpp) of the unreleased model.
OpenBMB has a MiniCPM5 MoE coming: MiniCPM5-15B-A2.5B, 15B total and 2.5B active. A ModelBest engineer opened a llama.cpp PR for it this morning, tested against GGUFs of the unreleased model. https://github.com/ggml-org/llama.cpp/pull/32386
OpenBMB develops the "MiniCPM" series of small models. When it released the text-recognition model MiniCPM-OCR 1.0 in October, support in the inference software vLLM had also been proposed five days before the weights were released.
PR adds inference support for "MiniCPM5 MoE"
According to the PR description, the change adds inference support for running MiniCPM5 MoE in the GGUF format. It lists the architecture as "gated GQA," a mechanism for the attention computation; a leading ordinary layer (dense FFN); and experts selected per input (routed experts) along with shared experts. It reuses the existing MiniCPM5 tokenizer (the part that splits text into tokens), and conversion from the original format to GGUF is outside the PR's scope.
The model name MiniCPM5-15B-A2.5B appears in the PR's description of its tests. The figures of 15 billion total and 2.5 billion active parameters, and the word "unreleased," do not appear in the PR description and come only from the bot.
Author states tests are "not model-quality evaluation"
The author writes that pre-existing MiniCPM5-15B-A2.5B GGUFs were tested on a single NVIDIA "H800" GPU (80GB). Ordinary generation passed in three formats: 16-bit BF16, 8-bit Q8_0 and Q4_K_M, a mixed-precision format centered on 4 bits. Generation also passed with DSpark, an implementation of "speculative decoding," a technique for speeding up generation, using a file named "MiniCPM5-15B-A2.5B-1025-DSpark.gguf."
The author also stated that the checks looked at whether inference runs and does not crash, and are not a model-quality evaluation or performance benchmark. No speedup claim is made. Ordinary generation and DSpark showed token-level differences in output, the cause has not been fully characterized, and output equivalence has not yet been established, the author said.
Author's affiliation and release status unconfirmed
The bot described the PR's author as "A ModelBest engineer," but the public profile of the author's GitHub account, "cyxu0401," has no name, company or bio, and the affiliation cannot be confirmed. In the PR's records, the author's association with the llama.cpp repository is listed as "NONE."
CISC, a llama.cpp member, pointed out in a PR comment on the evening of Nov. 16 Japan time that conversion code is required. In response to the statement that pre-existing GGUFs were tested, CISC wrote "Not public yet AFAICT?" Collaborator ngxson then asked tc-mb to confirm. tc-mb's GitHub bio reads "Member@MiniCPM-V & MiniCPM-o," but no company is listed.
The Hugging Face page for openbmb/MiniCPM5-15B-A2.5B could not be confirmed as a public model when ORIENS retrieved it early on Nov. 17. No announcement of the model could be found among posts on OpenBMB's official X account since Oct. 1 either.
- Draft PR for MiniCPM5 MoE support submitted to llama.cppLeak
- Reviewing member comments that the model may not be public yetThird party
- Detection bot @aitrackerbot reports MiniCPM5-15B-A2.5BLeak





