
MiniCPM5-15B-A2.5B: Ported by Four Runtimes, Downloadable by None
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 189 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 204 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 132 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 77 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 93 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 211 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
On 2026-11-16 a pull request appeared in llama.cpp titled "xcy/models: add MiniCPM5 MoE inference support". It is a careful piece of work — a new architecture entry, a conversion path, tests — and it carried one line in its validation section that a maintainer stopped on: the author had run it against "pre-existing MiniCPM5-15B-A2.5B GGUFs" on an H800. CISC asked the obvious question. Not public yet AFAICT? tc-mb, whose GitHub profile lists membership in the MiniCPM-V and MiniCPM-o projects, answered in one word: "ok". The pull request was closed unmerged two days later, and the model those GGUFs belong to — MiniCPM5-15B-A2.5B, the sparse mixture-of-experts entry in OpenBMB's MiniCPM5 line — still has no repository, no model card, no licence file and no announcement we could find anywhere.
So this is a what-we-know-so-far piece, and the shape of the evidence matters more than usual. There is no product page to read and no vendor benchmark to argue with. What exists is a set of engine patches from four separate inference stacks, written by people who had the checkpoint in front of them, plus their test logs. That is enough to describe MiniCPM5 MoE's architecture fairly precisely and to say something honest about how mature the support is. It is not enough to tell you what the model scores, what it costs, or when you can run it.
What the name is allowed to tell you
The checkpoint name is the only spec sheet anyone outside ModelBest has published. MiniCPM5-15B-A2.5B reads as 15B total parameters with 2.5B active per token — the total-and-active convention every sparse MoE in this size class since Mixtral has used. Nothing in any patch contradicts it, and the SGLang port's own test notes refer to "the original 15B checkpoint config" resolving without modification, which is consistent.
Everything past that is guesswork, and it is worth being blunt about why. The SGLang patch adds a MiniCPM5Config class whose defaults read 32 layers, 4,096 hidden, 32 heads, 64 routed experts, 2 shared experts and a 1,407-wide expert FFN. Those are class defaults inherited from a GLM-4 template, not the checkpoint's values — the same file defaults to a 32,000-token vocabulary, which no 2026 frontier-adjacent model has shipped with. Do not read them as specs. The one number that looks like it might be real is the routed scaling factor of 3.66, which appears in the llama.cpp architecture test and in the conversion test fixture; a fixture is still a fixture, so treat 3.66 as "the number the tests were written around" rather than as a confirmed constant.
Architecture, as read off four patches
Four independent ports converge on the same description, which is the most reassuring thing about this leak. The SGLang patch is the most explicit: the model "follows GLM-4.5 MoE, with three differences" — gated attention, full rotary embeddings without q/k normalisation, and a router computed in the model's own dtype rather than upcast.
• Gated attention — the query projection emits a per-head sigmoid gate alongside the query, interleaved inside each head's rows. The llama.cpp conversion has to split them apart tensor by tensor, which is why both ports call it out.
• Dense lead-in, then experts — the first first_k_dense_replace layers are ordinary FFNs; only the layers after that route. tc-mb's fork patch notes that neither the existing qwen35moe nor minicpm architectures model that shape, which is why MiniCPM5 MoE needed its own architecture entry rather than a flag on an existing one.
• Routed plus shared experts — sigmoid gating, normalised top-k weights, and a shared-expert branch alongside the routed ones. llama.cpp reads an FP32 router correction bias (e_score_correction_bias) per MoE layer, the same bias term DeepSeek-style routers use to keep expert load balanced.
• Full rotary, no q/k norm — a plain RoPE across the head dimension, with the conversion preserving head dimension explicitly rather than deriving it from hidden size divided by heads.

DSpark is the second half of the story, and it is the reason this leak is interesting rather than merely tidy. DSpark is speculative decoding with a trained draft model — the same algorithm OpenBMB already shipped for MiniCPM5-2B and MiniCPM-V 4.7 — and both the SGLang and llama.cpp patches add MiniCPM5 MoE as a DSpark target. A separate draft checkpoint named MiniCPM5-15B-A2.5B-DSpark is referenced in the llama.cpp docs example, which means the draft exists too, in whatever private store the target lives in.
Four engines, one unreleased checkpoint
The timeline is what makes this worth writing about. Ports do not normally precede a release by this much across this many stacks:
• 2026-09-17 — Megatron-LM patches on a MUSA fork enable MoE capacity-factor handling and dropless routing as the default for "minicpm5". Training-side, and the earliest trace we found.
• 2026-10-14 — InfiniTensor's InfiniLM opens a patch enabling MiniCPM5 MoE on the Denglin QY backend with FlashAttention and long-context MoE/RoPE kernels; a follow-up on 2026-11-12 adds CUDA-graph capture safety and a bf16 decode path.
• 2026-11-13 — tc-mb's llama.cpp-omni fork adds the MiniCPM5 MoE architecture, and a second patch the same day wires expert regions into the quantizer. The verification section reports The capital of France is completing to Paris. and 17×23 returning 391.
• 2026-11-16 — the upstream llama.cpp pull request. Closed 2026-11-18, unmerged, with the conversion path pushed out of scope and revisited three days later in the fork.
• 2026-11-19 — SGLang opens a port at 09:39, closes it, and reopens it at 10:19 under a renamed branch. The same morning, an Intel engineer opens an optimum-intel patch whose very first code block is an export command naming openbmb/MiniCPM5-15B-A2.5B as its input.

That last one is the closest thing to a smoking gun. The optimum-intel patch does not describe a hypothetical model — it runs optimum-cli export openvino -m openbmb/MiniCPM5-15B-A2.5B --weight-format int4 and reports accuracy against the original. Whoever wrote it had the weights. The Hub path it names returns nothing to an anonymous request, and OpenBMB's own MiniCPM README still lists 2026-09-07 and the 2B as the latest release, with no MoE entry anywhere in its model table.
What the test logs actually show
This is the part to read slowly, because the numbers in these patches are engineering checks, not benchmarks, and two of them are failures.
The llama.cpp run is the broadest. BF16, Q8_0 and Q4_K_M weights all produced and accepted DSpark draft tokens on an H800 with the draft at --spec-draft-n-max 2. Eight tokenizer samples matched the Hugging Face tokenizer exactly. The CPU-versus-CUDA NMSE on the synthetic architecture test was 1.00e-06. And then the caveat, in the author's own words: greedy ordinary and DSpark decoding "showed token-level differences", the cause "has not been fully characterized", and output equivalence "is not established". There is no logits-parity run against the original checkpoint, no quality evaluation and no performance number. The author states plainly that no speedup claim is made.
The SGLang run is more informative because it isolates a variable. With Triton attention, DSpark executed 93 verification steps over four 64-token generations and diverged from ordinary decoding on every prompt. With FA3 attention, the same 15B checkpoint and draft produced matching greedy token IDs on all four. The Triton discrepancy is still open in the patch as written. Read that as what it is: a kernel-path difference someone found and has not yet explained, not a statement about the model.
The OpenVINO numbers are the only quantitative quality signal in the whole set, and they are modest. Against the original model in FP32 across 26 prompts and 128 new tokens, the WWB text-similarity score was 0.9934 for bf16 on CPU, 0.9455 for int8 and 0.8710 for int4 under the GenAI SDPA backend; on a B70 GPU, 0.9498 for int8 and 0.8646 for int4. An int4 conversion landing at 0.87 is a conversion-fidelity result — it says the quantised graph drifts from the reference, not that the reference is good. Nobody has published a benchmark score for MiniCPM5-15B-A2.5B, and until someone does, there is no basis for calling it strong or weak.

Where it would sit if it ships
OpenBMB's MiniCPM5 line is, so far, an argument about density. The 1B shipped 2026-05-19 and the 2B on 2026-09-07, both dense, both Apache-2.0, both built for on-device deployment — the 2B is a plain LlamaForCausalLM with no custom kernels, which is precisely why it ported to llama.cpp, MLX, ONNX and half a dozen NPU toolchains within weeks. A 15B-total, 2.5B-active sibling would be the first MiniCPM5 that is not an edge model: the same active budget as the 2B, roughly six times the stored weights, and a routing layer that has to be served well to be worth its memory.
That framing also explains the strange shape of this leak. If MiniCPM5 MoE is aimed at serving rather than phones, then the engine ports are not a nice-to-have shipped after launch — they are the launch prerequisite, and somebody is doing them early. The fact that the same GitHub account opened both the llama.cpp and the SGLang ports, that a MiniCPM project member is landing the fork patches, and that a draft checkpoint already exists for DSpark all point the same way: this is a model being prepared for release by people with access to it, not a rumour assembled from changelogs.
What would change this picture
Three things, in order of how much they would settle.
First, a repository. A Hub or ModelScope page with a config, a licence and a card would convert most of the paragraph above from inference to fact, and would immediately give the config defaults I warned about something to be checked against.
Second, a primary-source benchmark. The MiniCPM5-2B card carried a 53.9 suite average and per-task cells for LiveCodeBench, AIME and SWE-bench; a 15B-A2.5B with no equivalent table is a model whose whole value proposition is unmeasured. An Artificial Analysis entry would do it, since that is independent of the vendor.
Third, a merged port. Every patch described here is either closed, draft, or sitting on a private fork. Support that exists only in an unmerged pull request is support you cannot install, and — as the Triton-versus-FA3 divergence shows — correctness on this architecture is still being established kernel by kernel.
The practical consequence for a reader today is simple: nothing to download, nothing to call, nothing to migrate. What a model like this changes when it does land is the decision about where a workload runs, and that decision is usually made under time pressure with incomplete evidence — which is the situation a routing layer exists for. OrcaRouter puts 200+ hosted models behind one API at provider list price with no markup added, so the comparison you would want to run, this checkpoint against whatever you call today, costs nothing to stand up once someone hosts it; automatic failover means a first-week model never has to hold a production path on its own. None of that requires MiniCPM5-15B-A2.5B to be available, and we do not serve it — it is the safety net around the models you already depend on while an unreleased one proves itself.
Until then the honest summary is the one the llama.cpp thread gave us: the GGUFs exist, and they are not public. Four runtimes have already written the code to load them. Watch the OpenBMB organisation page and the two open pull requests, in that order — a repository appearing is the signal, and a merged port is the confirmation that the architecture is what these patches say it is.
