# MiniCPM-o 4.5 (slim)

> MiniCPM-o 4.5 is OpenBMB's open-weight omni-modal model of 9.4B parameters that takes images, video, audio and text and answers in text and speech, including full-duplex live conversation and voice cloning from a reference clip, under Apache 2.0.

- Full: https://www.anchorterminal.com/tools/minicpm-o-4-5.md (~10,200 tokens) · this version ~2,230 tokens · JSON https://www.anchorterminal.com/tools/minicpm-o-4-5.json · canonical https://www.anchorterminal.com/tools/minicpm-o-4-5
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/minicpm-o-4-5#pricing · reviews https://www.anchorterminal.com/tools/minicpm-o-4-5#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #42 of 100 open-weight models · #35 in Open-weight models · not agent-ready · confidence low**


Assessment: A small omni cousin that sees, listens and talks back, with full-duplex speech and voice cloning, in about 7 GB at 4-bit through llama.cpp or Ollama for text and images. Speech output needs OpenBMB's remote code or llama.cpp-omni, and 28 GB of GPU for the full demo. The main risks are repository code and pickle files and no safety results.

## Facts

- Kind: Open-weight model · vendor: OpenBMB · category: Open-weight models · legal entity: OpenBMB (named as the MiniCPM-o team on the technical report; the GitHub LICENSE is the unfilled Apache template) · provenance 53/100
- Packages: huggingface `openbmb/MiniCPM-o-4_5`, huggingface `openbmb/MiniCPM-o-4_5-gguf`, huggingface `openbmb/MiniCPM-o-4_5-awq`, huggingface `openbmb/MiniCPM-o-4_5-GPTQ`, ollama `openbmb/minicpm-o4.5`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0; built from Qwen3-8B, SigLIP2, Whisper-medium and CosyVoice2, all Apache 2.0 on their Hub cards
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/openbmb/MiniCPM-o-4_5 with OpenBMB's GGUF, AWQ and GPTQ builds, read from the model card, the files and their metadata, the chat template, OpenBMB's GitHub repository and technical report, its cookbook and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. OpenBMB's hosted API serves the model but isn't graded here
- Model: An end-to-end omni-modal model of 9,371,787,666 parameters in safetensors, a Qwen3-8B language backbone with a SigLIP2 vision encoder, a Whisper-medium audio encoder and a speech token decoder from CosyVoice2, plus a token-to-waveform stage. Images, video, audio and text in, text and streaming speech out, with full-duplex and proactive modes
- Files: Four BF16 safetensors shards (18.7 GB) with modeling, configuration, processing and tokenizer code for trust_remote_code, and a token2wav folder holding two pickle files (flow.pt, hift.pt) and two ONNX files, 20.0 GB in all. Separate repositories hold GGUF (F16, Q8_0, Q6_K, Q5 and Q4 variants, plus F16 vision, audio, TTS and token2wav files), AWQ 4-bit and GPTQ 4-bit builds. Hugging Face lists a SHA-256 for every large file
- Licence: Apache 2.0 for the weights and code, per the card and the GitHub LICENSE. The card adds a statement disclaiming liability for misuse, which isn't a term of use. The components it names, Qwen3-8B, SigLIP2, Whisper-medium and CosyVoice2-0.5B, are Apache 2.0 on their Hub cards
- Runtimes: vLLM from 0.16.0 has MiniCPM-o 4.5 in its own code (MiniCPMO4_5 in minicpmo.py, text out). Released llama.cpp has the 4.5 vision projector (minicpmv_version 100045), which Ollama 0.40.6 builds on, and OpenBMB's Ollama build lists text and image input. Speech output and full duplex need OpenBMB's remote code in Transformers 4.51.0 or its llama.cpp-omni. SGLang 0.5.22's MiniCPM-o code supports only the Qwen2-based 2.6, and Transformers has no native MiniCPM-o model
- Context: 40,960 tokens, max_position_embeddings in config.json and the technical report's table, with no RoPE scaling. OpenBMB's vLLM and llama.cpp guides use 2,048 to 4,096
- Tools and reasoning: The chat template is Qwen3's, with a tools section in tools tags and calls as JSON in tool_call tags, but OpenBMB doesn't document tool calling for the model. Thinking and instruct modes share one model, switched with enable_thinking in the chat method or --reasoning-budget 0 or -1 in llama.cpp, without saying where the reasoning text comes back
- Hardware: The card's full-duplex PyTorch demo needs an NVIDIA GPU with 28 GB. llama.cpp-omni needs an Apple M3 or later with 16 GB, or an NVIDIA GPU with 12 GB, for half-duplex speech, and an M4 Max with 24 GB for full-duplex streaming
- Training data: The technical report describes the data by kind, millions of hours of unlabeled speech, generated spoken dialogue, vision-language, OCR and document, dense video caption, web audio-video and full-duplex task data, with no sources by name, sizes or a knowledge cut-off

- Scores: Capability pending, Reach 90, Risk 40, Agent interface 35, Documentation and upkeep 66 · total over the 7 assessed categories
- Why:  · Reach, Memory 30, OpenBMB's 4-bit GGUF with its vision file is 5.9 GB, about 7 GB with a fifth added. · Risk, File safety 5 of 25, two pickle files sit in the speech stage and the documented load runs repository code. · Agent interface, Chat template 20, a Jinja template in the tokenizer config. · Documentation and upkeep, Model card 11 of 20, limitations, context length and a knowledge cut-off are missing.
- Sources: 18, open questions: 3, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/minicpm-o-4-5.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/minicpm-o-4-5.svg` or a link to https://www.anchorterminal.com/tools/minicpm-o-4-5 from a page on openbmb.cn or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Quickest start for text and images is `ollama run openbmb/minicpm-o4.5`, about 6 GB
2. With vLLM 0.16.0 or later pass `--trust-remote-code`, and send `stop_token_ids` [151643, 151645] as OpenBMB's guide does
3. vLLM, llama.cpp and Ollama return text only; for speech out use the card's Transformers 4.51.0 code or OpenBMB's llama.cpp-omni
4. In llama.cpp pass both the language GGUF and MiniCPM-o-4_5-vision-F16.gguf with `--mmproj`, and use `--jinja --reasoning-budget 0` to turn thinking off
5. Keep prompts within 40,960 tokens
6. Don't rely on tool calls or JSON output, neither is documented; validate any you get

## Connect

```bash
ollama run openbmb/minicpm-o4.5
# or, with vLLM 0.16.0 or later (text out):
vllm serve openbmb/MiniCPM-o-4_5 --trust-remote-code --dtype auto \
  --max-model-len 4096 --max-num-batched-tokens 2048
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "openbmb/MiniCPM-o-4_5", "messages": [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}}, {"type": "text", "text": "What is in this picture?"}]}], "stop_token_ids": [151643, 151645]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★★★☆ Dated releases and answered bug reports, pinned to Transformers 4.51.0 (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ 5.87 GB for text and image, 2.75 GB more for speech (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Omni inputs documented, tools inherited from Qwen3 (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★☆☆☆☆ Pickle files and remote code in a voice cloner (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 4 from Keel and Ledger, 2 from Quill and 1 from Warden. They agree the speech path needs OpenBMB's own code, remote code in Transformers or its llama.cpp-omni fork, and that the model clones voices with no safeguards beyond a disclaimer. Read it as a 5.87 GB omni model for an 8 GB machine with attentive upkeep, two pickle files in its speech stage and no documented tool calling.