MiniCPM5-2B: a 2.5B model that actually beats its size class, and loads today
OpenBMB's MiniCPM5-2B shipped September 7 as a plain Llama-architecture checkpoint — no fork, no waiting on llama.cpp support — and it is small enough to sit comfortably on a 6 GB iPhone.
OpenBMB released MiniCPM5-2B on September 7. It is a 2.52 billion parameter dense model (1.98B if you don't count the embedding table), Apache 2.0, with GGUF, MLX and GPTQ builds up alongside the raw weights from day one. The part that matters more than the benchmark table: it uses a standard LlamaForCausalLM architecture — 42 layers, grouped-query attention with 16 query heads and 2 key/value heads. That means llama.cpp, and anything built on it, reads the file with no custom kernel and no waiting on an upstream PR to merge. That is not a given — we wrote about K2 Horizon here two days ago precisely because its 7B model shipped GGUF files that llama.cpp still can't open. MiniCPM5-2B has none of that problem.
What it actually is
MiniCPM5 is the fifth generation of a model line OpenBMB (a Tsinghua-affiliated lab) has been shipping since 2024, built specifically for phones and other memory-constrained devices rather than retrofitted onto one after the fact. This release is one checkpoint: 2B-class, dense, 131,072-token native context. A smaller MiniCPM5-1B came out back in May; this is not that model, and not a resize of it.
The GGUF repository lists a Q4_K_M build at 1.56 GB and a Q8_0 build at 2.68 GB. Both are small by the standards of anything else we've covered on this blog.
Does it fit on your phone
Using the RAM arithmetic from our earlier post on iPhone memory — roughly 44% of physical RAM survives to become usable weight budget, after iOS's jetsam behavior and runtime overhead are accounted for:
| Device | Weight ceiling | Q4_K_M (1.56 GB) | Q8_0 (2.68 GB) |
|---|---|---|---|
| 4 GB — iPhone 12/13 | 1.29 GB | ✗ too big | ✗ too big |
| 6 GB — iPhone 12–14 Pro, 14, 15 | 2.23 GB | ✓ fits | ✗ too big |
| 8 GB — iPhone 15 Pro, 16, 16e, 17 | 3.18 GB | ✓ fits | ✓ fits |
So Q4_K_M is the size that matters here: it clears the 6 GB ceiling with room to spare, which is the first phone class most people are actually still carrying. It does not fit a 4 GB iPhone 12 or 13 — those phones are already spoken for by 1B-class models, and nothing about this release changes that.
The benchmark claims, and how much to trust them
OpenBMB reports an average of 53.9 across 34 benchmarks, which they call SOTA for the 2B class, against a published table that includes Qwen3.5-2B (33.2 average, same table) and Gemma-4-E2B-it. That is the lab's own comparison, so treat the exact gap with the usual skepticism you'd apply to any vendor's own chart. What's harder to wave off is Artificial Analysis's independent Intelligence Index, where MiniCPM5-2B scored 23 — first place among every open model under 4B parameters they track, ahead of the field rather than just ahead of two models OpenBMB chose to put in a table.
On individual benchmarks OpenBMB lists: 69.1 on LiveCodeBench v6, 86.5 on AIME 2026, 94.6 on MATH-500, 66.6 on BFCL v4 (tool use), 97.1 on tau2-Bench Telecom (agentic tool-calling). Those are code, math, and tool-use numbers, and they're where the model claims its biggest edge — OpenBMB's own writeup singles out coding, math and agentic tasks as the strengths, and concedes ground to Qwen3.5-2B on instruction-following benchmarks like IFEval. That tracks with what these training recipes typically trade off: heavy RL and on-policy distillation on reasoning-shaped tasks tends to cost you some polish on "just follow the instruction" evals.
We have not run these benchmarks ourselves, and this network's access to Hugging Face was blocked while researching this post, so the file sizes and benchmark numbers above come from OpenBMB's own GitHub README and from independent write-ups (Artificial Analysis, MarkTechPost) that were cross-checked against each other rather than read directly off the model card. If a number here turns out to be off, that's the reason — flag it and we'll fix it.
Should you download it
If you're on a 6 GB-or-better iPhone and want a genuinely small model that leans toward code, math and tool-calling rather than general chat, yes — this is worth the 1.56 GB. It is the rare small release that ships a working llama.cpp path on day one instead of a promise that one is coming, and the independent Intelligence Index placement is real signal, not just a lab grading its own homework. If you're on a 4 GB iPhone, this one still isn't for you — that ceiling hasn't moved, and won't until something meaningfully smaller than 1B shows up with numbers to back it.