MiniCPM5-1B — LiteRT-LM (on-device)

Main Model Card: openbmb/MiniCPM5-1B

MiniCPM5-1B packaged in the .litertlm format so it runs fully on-device — Android, iOS, Desktop and IoT — with the LiteRT-LM framework. No internet connection required at inference time.

👉 Not sure which file to pick? Download MiniCPM5-1B.litertlm

That's the one and only file here, and it's meant to "just work". You don't need to understand quantization to use it.

File	What it is	Who it's for
MiniCPM5-1B.litertlm	The ready-to-run model (INT8, ~1.1 GB)	Everyone — this is the default
chat_template.jinja	The chat format (auto-used by the runtime)	Advanced / reference

Try it in one command (Desktop)

uv tool install litert-lm
uvx litert-lm run --from-huggingface-repo=luispoveda93/MiniCPM5-1B MiniCPM5-1B.litertlm --prompt="What is the capital of France?"

Run it on your phone (Android)

1. Install the Google AI Edge Gallery app.
2. Open Model manager → “+” → Import from HF, paste this repo's link, and download MiniCPM5-1B.litertlm.

See the Edge Gallery wiki for details.

About this model

MiniCPM5-1B is a compact 1B decoder-only model from OpenBMB (standard LlamaForCausalLM), built for local assistants, coding agents, and tool use, with a hybrid <think> chat mode and long-context support.

Item	Value
Parameters	1,080,632,832 (~1B)
Layers	24
Attention (GQA)	16 query heads / 2 KV heads
Context length	131,072
Quantization	INT8 dynamic-range (weights), FP32 activations
Signatures	prefill_128, prefill_512, decode (KV-cache 1280)

Recommended sampling (from the source model card)

Mode	Params	Enable
Think	temperature=0.9, top_p=0.95	enable_thinking=True
No Think	temperature=0.7, top_p=0.95	enable_thinking=False

How this was built (transparency)

Converted from the BF16 source (openbmb/MiniCPM5-1B) with Google AI Edge tooling:

• Re-authoring + TFLite conversion: ai-edge-torch Generative API. MiniCPM5-1B maps onto the standard decoder-only building blocks (hidden 1536, 24 layers, GQA 16/2, head_dim 128, RMSNorm 1e-6, RoPE θ=5e6, untied lm_head, vocab 130560).
• Quantization: INT8 dynamic-range weight-only.
• Packaging: litert-lm-builder, native minicpm5 runtime type, source Hugging Face tokenizer, start token <s>, stop tokens [1, 130073] (</s>, <|im_end|>), thought reasoning channel, and an ExecutorMetadata section declaring all 48 KV-cache state buffers (required for on-device prefill).
• Verification: the re-authored model was checked against the original transformers implementation — last-token logits matched with cosine 1.0000, identical argmax, 5/5 top-5 overlap; parameter count matches exactly. The packaged bundle was also run end-to-end with litert-lm run and generates correctly (reasoning channel parses as expected).

This is a personal community contribution, not an official OpenBMB or Google release. An INT4 (wi4b32) variant with a smaller footprint is available from the official litert-community/MiniCPM5-1B repo.

Community-Contributed Evaluation: Spanish, Catalan, Basque, Galician (0-shot)

⚠️ Note: This section is a community addition — it does not appear in the original openbmb/MiniCPM5-1B model card. It was benchmarked independently by this repo's maintainer and is provided as supplementary information, not an official OpenBMB result.

To understand how this compact 1B EN/ZH-pretrained model performs on Spain's co-official languages, we ran the exact task selection published on the Salamandra-2b-instruct model card — a 2.25B model whose pretraining specifically oversampled these languages 2x — using the same LM Evaluation Harness task suite (SpanishBench, CatalanBench, BasqueBench, GalicianBench), 0-shot, chat template applied. Evaluation was run on the BF16 source model (openbmb/MiniCPM5-1B), not the on-device .litertlm build in this repo. The Salamandra-2B column is transcribed verbatim from its published card and was not re-run.

Spanish

Category	Task	Metric	MiniCPM5-1B	Salamandra-2B
Commonsense Reasoning	xstorycloze_es	acc	51.69	61.95
NLI	wnli_es	acc	43.66	49.30
	xnli_es	acc	42.49	48.52
Paraphrasing	paws_es	acc	57.30	57.10
QA	xquad_es	f1†	0.00	29.60
Translation	flores_es	bleu	0.00	15.92

^† Salamandra's card lists this task under metric "acc"; the installed xquad_es task only computes f1/exact_match, so f1 is reported here for MiniCPM5-1B.

Catalan

Category	Task	Metric	MiniCPM5-1B	Salamandra-2B
Commonsense Reasoning	copa_ca	acc	50.80	71.00
	xstorycloze_ca	acc	48.64	61.61
NLI	wnli_ca	acc	54.93	56.34
	xnli_ca	acc	37.83	52.54
Paraphrasing	parafraseja	acc	56.75	64.48
	paws_ca	acc	56.45	60.30
QA	arc_ca_easy	acc	28.87	52.15
	arc_ca_challenge	acc	20.22	30.03
	openbookqa_ca	acc	21.80	29.20
	piqa_ca	acc	53.32	63.93
	siqa_ca	acc	33.76	42.32
Translation	flores_ca	bleu	0.00	20.04

Basque

Category	Task	Metric	MiniCPM5-1B	Salamandra-2B
Commonsense Reasoning	xcopa_eu	acc	48.20	55.20
	xstorycloze_eu	acc	49.77	55.26
NLI	wnli_eu	acc	52.11	52.11
	xnli_eu	acc	33.91	45.89
QA	eus_exams‡	acc	24.40	24.94
	eus_proficiency	acc	24.28	25.15
	eus_trivia	acc	30.44	27.35
Reading Comprehension	eus_reading	acc	25.85	28.69
Translation	flores_eu	bleu	0.00	7.93

^‡ eus_exams is a group of 29 exam sub-tasks; MiniCPM5-1B's score is the sample-count-weighted average across all sub-tasks.

Galician

Category	Task	Metric	MiniCPM5-1B	Salamandra-2B
Paraphrasing	parafrases_gl	acc	52.72	53.40
	paws_gl	acc	53.95	56.85
QA	openbookqa_gl	acc	21.60	26.20
Translation	flores_gl	bleu	0.00	17.53

Bold values indicate MiniCPM5-1B matches or exceeds the Salamandra-2B-instruct published baseline despite the parameter and pretraining-data gap. MiniCPM5-1B has no targeted Spanish/Catalan/Basque/Galician pretraining data; the gaps on Basque and Galician (and all FLORES/BLEU translation tasks, which scored ≈0 across every language) are the expected result of that mismatch, not a defect in the on-device conversion. This comparison is offered as context for anyone considering MiniCPM5-1B for these languages — it is not a claim of official parity with Salamandra.

License

Apache-2.0, following the source model.