Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › MiniCPM5 family

MiniCPM5-2B

Hugging Face ↗ Follow changes

openbmb Apache 2.0 out 49 days ago

2.52B parameters, text in, text out, 128K tokens of context.

Verified 2026-10-25 Rank #297 in LLMs · score 84 method v1 Sources 3 Downloads 30d 1.2M · #98 of 619 tracked

Changed this week: 6 new quantized builds on the Hub (111 in all).2026-10-24 · 5 new quantized builds on the Hub (110 in all).2026-10-24

Can I run it?1.5GB at INT4Estimatefits a 16 GB cardmethodCommunity: MiniCPM5 2B: SOTA or Scam? Let's Test Locally - Frontier Models →
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?1.2Mdownloads, 30 days#98 of 619 we track · top 16%+27 likes, downloads −5% this week1,813 likesHugging Face2026-10-24

Worth reading

Picked by rule from 10 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Latest · 43 days ago

OpenBMB MiniCPM5-2B

OpenBMB released MiniCPM5-2B on ~2026-09-07 as a 2.5B dense model achieving the highest AA Intelligence Index v4.2 score of 15 among open-weight models under 4B parameters.

  • The model uses a hybrid think/no-think architecture combined with RL and on-policy-distillation UltraData training.
  • It is the first sub-3B model to demonstrate meaningful repo-level agent coding capabilities.
kinonn-bot.github.io 2026-09-13 674 words

10 best of 10 tracked

The 3 picks above, plus these 7. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. ArticleMiniCPM5-2B – new leading ≤4B modelartificialanalysis.ai · 2026-09-07 · 4 HN pts
  2. CodeDesk-Pet – a local-first desktop pet powered by MiniCPM5github.com · 2026-07-07 · 2 HN pts
  3. ArticleMiniCPM5-2B: Does the 2B "SOTA" Claim Survive Real Testing?mindstudio.ai · undated
  4. ArticleMiniCPM5-2B: how 2 KV heads pay for a 131K windowai.thesatyajit.com · undated
  5. ArticleMiniCPM5-2B — slm.expertslm.expert · undated
  6. ArticleMiniCPM5-2B review: the 2B model that beats bigger oneseesel.ai · undated
  7. ArticleMiniCPM5-2B review: how good is this tiny local LLM, really?creatorprivacykit.com · undated

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 111 quantized builds and 25 Spaces.

  • As a service0no API provider listed
  • On own hardware111quantized builds
  • Built on11479 fine-tunes · 34 adapters · 1 merges
  • In apps25Hugging Face Spaces

Top Spaces running it: victor/MiniCPM5-2B-WebGPU-Pi ♥ 102 · openbmb/MiniCPM5-2B-Demo ♥ 68 · ProCreations/minicpm5-2b-webgpu ♥ 8

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 1.5 GB at INT4, 3.0 GB at FP8, 6.0 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.5 GB of weights (estimate; KV cache and activations extra)INT41.5 GBFP8: about 3.0 GB of weights (estimate; KV cache and activations extra)FP83.0 GBBF16: about 6.0 GB of weights (estimate; KV cache and activations extra)BF166.0 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 1.5 GB of weights (estimate; KV cache and activations extra)INT41.5 GBFP8: about 3.0 GB of weights (estimate; KV cache and activations extra)FP83.0 GBBF16: about 6.0 GB of weights (estimate; KV cache and activations extra)BF166.0 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2026-09-06), amber triangles are 5 dated articles (each links to it), grey ticks are repo commits.

AugSepOct2026-09-29: 1 repo commit2026-09-12: 2 repo commits2026-09-10: 2 repo commits2026-09-09: 2 repo commits2026-09-08: 2 repo commits2026-09-07: 4 repo commits2026-09-06: 4 repo commitsMiniCPM5-2B – new leading ≤4B model · 2026-09-07Desk-Pet – a local-first desktop pet powered by MiniCPM5 · 2026-07-07OpenBMB MiniCPM5-2B · 2026-09-13MiniCPM5-2B: The Best Sub-Agent Model Yet? - Frontier Models · 2026-09-10MiniCPM5 2B: SOTA or Scam? Let's Test Locally - Frontier Models · 2026-09-07Weights 2026-09-06 · 2026-09-06Weights 2026-09-06now
AugSepOct2026-09-29: 1 repo commit2026-09-12: 2 repo commits2026-09-10: 2 repo commits2026-09-09: 2 repo commits2026-09-08: 2 repo commits2026-09-07: 4 repo commits2026-09-06: 4 repo commitsMiniCPM5-2B – new leading ≤4B model · 2026-09-07Desk-Pet – a local-first desktop pet powered by MiniCPM5 · 2026-07-07OpenBMB MiniCPM5-2B · 2026-09-13MiniCPM5-2B: The Best Sub-Agent Model Yet? - Frontier Models · 2026-09-10MiniCPM5 2B: SOTA or Scam? Let's Test Locally - Frontier Models · 2026-09-07Weights 2026-09-06 · 2026-09-06Weights 2026-09-06now

What people built on it

225 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

225 derivatives on the Hub, led by 111 quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

111 quantized (Hugging Face model tree)79 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)34 adapters (Hugging Face model tree)111 quantized79 fine-tunes1 merge34 adapters
111 quantized (Hugging Face model tree)79 fine-tunes (Hugging Face model tree)1 merge (Hugging Face model tree)34 adapters (Hugging Face model tree)111 quantized79 fine-tunes1 merge34 adapters
GGUF Abiray/MiniCPM5-2B-heretic-abliterated-GGUF for llama.cpp, Ollama, LM Studio 35.7k downloads MLX mlx-community/MiniCPM5-2B-mlx-4Bit for Apple silicon 11k downloads NVFP4 Ttimms/MiniCPM5-2B-NVFP4 for vLLM, SGLang 2k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

What the card says

1 cited sentence

Summary generated, cited

MiniCPM5-2B is a 2.52B parameter dense Transformer model designed for on-device and local deployment.[1][2]

  1. parameters human
  2. “built for on-device, local deployment, and resource-const…”

Generated 2026-10-07 15:08 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

6 use cases

What it is good at 6 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesLlamaForCausalLM
Model typellama
Dtypebfloat16
Params by dtypeBF16: 2,516,756,480
Size bandM
MoEno
Context length131,072 (config.json max_position_embeddings)
Languagesen, zh
PapersarXiv:2506.07900, arXiv:2602.09003
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · LlamaForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · LlamaForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for MiniCPM5-2B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)6.0 GB · Single GPU (<24 GB)
FP8 (1 byte/param)3.0 GB · Edge (<4 GB)
INT4 (0.5 bytes/param)1.5 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Original model Family: openbmb/MiniCPM5 225 derivatives on the Hub

quantized: 111finetune: 79adapter: 34merge: 1

Top variants (40 shown, ranked by downloads)

Relation
Quant
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
Abiray/MiniCPM5-2B-heretic-abliterated-GGUF quantized GGUF — — apache-2.0 Abiray 42 35.7k
bartowski/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 bartowski 36 32.2k
NANI-Nithin/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 NANI-Nithin 8 17.2k
peculiar-ragdoll/Sharp-MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 peculiar-ragdoll 58 12.4k
mlx-community/MiniCPM5-2B-mlx-4Bit quantized mlx 2.52B M apache-2.0 mlx-community 3 11k
mradermacher/MiniCPM5-2B-i1-GGUF quantized GGUF — — apache-2.0 mradermacher 4 10.6k
abenzerps/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 abenzerps 11 8.7k
Abiray/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 Abiray 17 7.6k
mlx-community/MiniCPM5-2B-8bit quantized -8bit 2.52B M apache-2.0 mlx-community 9 6.7k
ngquocvinh/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 ngquocvinh 1 5.5k
saidutta69/MiniCPM5-2B-heretic quantized GGUF 2.52B M apache-2.0 saidutta69 0 5.2k
SC117/MiniCPM5-2B-abliterated-FIT-GGUF quantized GGUF — — apache-2.0 SC117 3 4.7k
lightonai/colgrep-agent-2B-GGUF quantized GGUF — — apache-2.0 lightonai 14 4.1k
abenzerps/MiniCPM5-2B-MLX-8bit quantized MLX 2.52B M apache-2.0 abenzerps 3 3.5k
dealignai/MiniCPM5-2B-CRACK-JANG_8M finetune — 2.52B M apache-2.0 dealignai 11 3k
mlx-community/MiniCPM5-2B-OptiQ-4bit quantized -4bit 2.52B M apache-2.0 mlx-community 4 3.1k
Mike0021/MiniCPM5-2B-ONNX quantized — — — apache-2.0 Mike0021 2 3.1k
nassimjp/MiniCPM5-2B-Pashto finetune — 2.52B M apache-2.0 nassimjp 0 3.1k
prithivMLmods/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 prithivMLmods 2 2.9k
huihui-ai/Huihui-MiniCPM5-2B-abliterated finetune — 2.52B M apache-2.0 huihui-ai 19 2k
infosave/MiniCPM5-2B-cmf quantized — — — apache-2.0 infosave 1 2.9k
gooseyai/MiniCPM5-2B-GGUF quantized GGUF — — cc0-1.0 gooseyai 0 2.9k
lightonai/colgrep-agent-2B finetune — 2.52B M apache-2.0 lightonai 34 842
logitlab/snap1-2b-GGUF quantized GGUF — — apache-2.0 logitlab 1 2.5k
julianmb/MiniCPM5-2B-NPU2 finetune — — — apache-2.0 julianmb 5 2k
WhiskyAKM/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 WhiskyAKM 1 2.1k
mradermacher/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 mradermacher 0 2.1k
ewin-reg/MiniCPM5-2B-RotSVDMix-Quantized quantized -Quant 2.72B M apache-2.0 ewin-reg 2 2k
jigs97022/minicpm5-2b-distilled-reasoning finetune — 2.52B M apache-2.0 jigs97022 5 1.8k
Ttimms/MiniCPM5-2B-NVFP4 quantized NVFP4 2.52B M apache-2.0 Ttimms 0 2k
KidIkaros/abliterated-minicpm5-2b-ggml quantized GGUF — — apache-2.0 KidIkaros 0 2k
nassimjp/MiniCPM5-2B-SFT-Pashto finetune — 2.52B M apache-2.0 nassimjp 0 2k
eaddario/MiniCPM5-2B-GGUF quantized GGUF — — apache-2.0 eaddario 2 1.8k
Atomic-Germ/MiniCPM5-2B-OpenNPU2 quantized GGUF — — apache-2.0 Atomic-Germ 0 1.9k
lee-loo/MiniCPM5-2B-MLX-bf16 finetune MLX 2.52B M apache-2.0 lee-loo 2 1.8k
mlboydaisuke/MiniCPM5-2B-CoreAI quantized — — — apache-2.0 mlboydaisuke 0 1.8k
mertkayacs/Tholos-2B finetune — 2.52B M apache-2.0 mertkayacs 2 1.6k
Johnny-Z/zako-pe finetune GGUF — — apache-2.0 Johnny-Z 14 999
abenzerps/MiniCPM5-2B-MLX-4bit quantized MLX 2.52B M apache-2.0 abenzerps 2 1.5k
Jeff98/MiniCPM5-2B-openvino finetune — — — apache-2.0 Jeff98 0 1.5k

Repository changelog

17 commits
changed in the last 7 daysOther: 2026-10-24, 2026-10-24all changes

Repository changelog 17 commits

DateCommitSHA
2026-09-29 Remove the dead `<tool_sep>` branch from chat_template.jinja (fixes OpenBMB/MiniCPM#379) (#19) f974000
2026-09-12 Enhance README with sampling parameter recommendations and Llama.cpp usage examples 12a3808
2026-09-12 Enhance README with sampling parameter recommendations and Llama.cpp usage examples 62b9b3b
2026-09-10 Add a LiteRT-LM row to the Deployment cookbook table abe115e
2026-09-10 Add a LiteRT-LM row to the Deployment cookbook table (#12) bca2410
2026-09-09 docs: update README cd199ce
2026-09-09 add `MiniCPM5-2B-DSpark-GGUF` d6c082b
2026-09-08 add the LiteRT-LM version of MiniCPM5-2B a063f08
2026-09-08 docs: update links of `Just RL II` and formatting in README files for clarity and consistency 0e9c66d
2026-09-07 update README 3497c46
2026-09-07 update README d3ad48d
2026-09-07 update README aac3680
2026-09-07 docs: update MiniCPM5-2B-GPTQ model link in README files b04e640
2026-09-06 docs: update training data and SGLang details d312cb6
2026-09-06 docs: update SGLang version requirement 85dfd5e
2026-09-06 upload MiniCPM5-2B ab8fd77
2026-09-06 initial commit 71b5e66

Articles and coverage

10 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Analysis10

MiniCPM5-2B – new leading ≤4B model
artificialanalysis.ai2026-09-07authorityHN 4 pts · 0 comments711 words
MiniCPM5-2B is an open weights model from OpenBMB that scores 12 on the Artificial Analysis Intelligence Index, significantly outperforming the median of 5 for its class. Released on September 7, 2026, it features a 131k token context window, 2.6 billion parameters, and an Apache 2.0 license. The model generated 74M output tokens during evaluation, which is higher than the median of 67M for comparable models.written, reviewed, cited
Desk-Pet – a local-first desktop pet powered by MiniCPM5
github.com2026-07-07authorityabout the MiniCPM5 familyHN 2 pts · 0 comments719 words
MiniCPM Desk Pet runs the MiniCPM5-1B-GGUF model locally on macOS Apple Silicon or Windows x64 systems requiring about 2 GB disk space. The application is distributed under the GNU AGPL-3.0-only license and supports downloading models from Hugging Face or ModelScope during the initial setup. It provides agent-aware reactions to coding tools and includes a neko-style persona adapter for desktop interaction.written, reviewed, cited
  • The app scans the machine for installed coding agents and prompts connection in one click.
  • Response speed depends on the user's chip, memory pressure, and the selected model.
  • Coding-agent reactions depend on each tool's integration behavior and may vary by version.
OpenBMB MiniCPM5-2B
kinonn-bot.github.io2026-09-13674 words
OpenBMB released MiniCPM5-2B on ~2026-09-07 as a 2.5B dense model achieving the highest AA Intelligence Index v4.2 score of 15 among open-weight models under 4B parameters. The model supports 131,072 tokens of native context and requires approximately 5.0 GB of VRAM in fp16, while leading benchmarks like SWE-bench Verified at 46.4 and LiveCodeBench v6 at 69.1.written, reviewed, cited
  • The model uses a hybrid think/no-think architecture combined with RL and on-policy-distillation UltraData training.
  • It is the first sub-3B model to demonstrate meaningful repo-level agent coding capabilities.
  • Deployment requires no custom kernels as it relies on standard LlamaForCausalLM implementations.
MiniCPM5-2B: The Best Sub-Agent Model Yet? - Frontier Models
frontiermodels.cc2026-09-10231 words
MiniCPM 2B (technically 2.5B) from OpenBMB claims to beat 4B-class models on function calling and agentic tasks. The model undergoes supervised fine-tuning on 400 billion deep-thinking tokens followed by a new reinforcement learning stage called RL2 that incorporates a critic model for improved training stability. Benchmarks show it edges out Qwen 3.5 4B on SWE-bench Verified while running at a fraction of the parameter count, though it struggles with complex visual tasks like SVG rendering.written, reviewed, cited
MiniCPM5 2B: SOTA or Scam? Let's Test Locally - Frontier Models
frontiermodels.cc2026-09-07228 words
MiniCPM5 2B undergoes local testing to validate lab benchmark claims against real-world tasks including an Oracle SQL partition-boundary bug, a question about the 1689 Treaty of Nerchinsk, and a creative front-end coding challenge requiring a self-contained animated HTML file. The model utilizes a three-stage training pipeline featuring reinforcement learning plus on-policy distillation (OPD) to achieve an 11-point gain on reasoning benchmarks. While factual accuracy is solid, response verbosity and slow reasoning chains remain practical drawbacks compared to competitors like Qwen.written, reviewed, cited
MiniCPM5-2B: Does the 2B "SOTA" Claim Survive Real Testing?
mindstudio.ai1,450 words
MiniCPM5-2B: how 2 KV heads pay for a 131K window
ai.thesatyajit.com6,879 words
MiniCPM5-2B — slm.expert
slm.expert
MiniCPM5-2B review: the 2B model that beats bigger ones
eesel.ai1,932 words
MiniCPM5-2B review: how good is this tiny local LLM, really?
creatorprivacykit.com1,773 words

Papers

2 papers

Papers 2

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • gemma-2-2b-it · 2.61B
    same size band (M), same task (text-generation), more providers
  • gemma-2-2b · 2.61B
    same size band (M), same task (text-generation), more providers
  • phi-2 · 2.78B
    same size band (M), same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:38 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:08 UTC · Last verified 2026-10-25

Follow MiniCPM5-2B

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.