Instructions to use efeisot/minicpm5-2b-rco with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use efeisot/minicpm5-2b-rco with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf efeisot/minicpm5-2b-rco
# Run inference directly in the terminal:
llama cli -hf efeisot/minicpm5-2b-rco

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf efeisot/minicpm5-2b-rco
# Run inference directly in the terminal:
llama cli -hf efeisot/minicpm5-2b-rco

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf efeisot/minicpm5-2b-rco
# Run inference directly in the terminal:
./build/bin/llama cli -hf efeisot/minicpm5-2b-rco

Use Docker

docker model run hf.co/efeisot/minicpm5-2b-rco

Llama new
LM Studio
Jan

vLLM

How to use efeisot/minicpm5-2b-rco with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "efeisot/minicpm5-2b-rco"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "efeisot/minicpm5-2b-rco",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/efeisot/minicpm5-2b-rco

Ollama
How to use efeisot/minicpm5-2b-rco with Ollama:
```
ollama run hf.co/efeisot/minicpm5-2b-rco
```
Unsloth Desktop

How to use efeisot/minicpm5-2b-rco with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf efeisot/minicpm5-2b-rco

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "efeisot/minicpm5-2b-rco"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use efeisot/minicpm5-2b-rco with Docker Model Runner:
```
docker model run hf.co/efeisot/minicpm5-2b-rco
```

Lemonade

How to use efeisot/minicpm5-2b-rco with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull efeisot/minicpm5-2b-rco

Run and chat with the model

lemonade run user.minicpm5-2b-rco-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use efeisot/minicpm5-2b-rco with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf efeisot/minicpm5-2b-rco

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default efeisot/minicpm5-2b-rco

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use efeisot/minicpm5-2b-rco with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf efeisot/minicpm5-2b-rco

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "efeisot/minicpm5-2b-rco" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

MiniCPM5-2B RCO GGUF (v2): code16, all-MLP4, down_proj-Q5

Curated mixed-precision quantization of openbmb/MiniCPM5-2B, packed as GGUF. The RCO tensors keep bit-exact RCO codes; the MLP weights are not dequantized and re-quantized through BF16.

This release supersedes the earlier experimental minicpm5-2b-rco-attn4-q6walls repository. It keeps only the variants that are useful as a size/quality trade-off.

code16, allmlp4, and downq5 contain non-standard GGML types RCO3 (44) and/or RCO4 (45). Stock llama.cpp cannot load them. Use the patched v0.7.0 runtime in runtime/ or apply runtime/llama.cpp-v0.7.0-rco-minicpm5.patch.

downq5 uses Q5_K for down_proj, so that tensor is standard K-quant; its gate_proj/up_proj MLP tensors are still bit-exact RCO4.

Which file?

Variant	Size	WikiText-2 PPL ↓	code2 p99 KL ↓	code2 top-1 ↑	MLP recipe	Role
`code16`	1.431 GB	14.9480	3.716	78.82%	66 x RCO3 + 228 x RCO4	Smallest
`allmlp4`	1.677 GB	14.3008	2.555	83.14%	126 x RCO4, no RCO3	Balanced
`downq5`	1.759 GB	14.0746	1.953	86.09%	84 x RCO4 + 42 x Q5_K down	Best RCO-lineage

allmlp4 removes every 3-bit MLP tensor. It is slightly smaller than the old mixed2a file and much better on held-out code. downq5 then upgrades all 42 down_proj tensors to Q5_K with the imatrix, which is the largest single quality jump in this release.

For comparison, the old mixed2a was 1.693 GB with code2 p99 3.507 and top-1 81.20%; allmlp4 is smaller and better. The old attn4 and mixed2a files remain available in the archived repository, but are not part of this curated release.

Sharp Q4_K_XL is still ahead on code tail: 1.596 GB, p99 1.237, top-1 89.31% on the same harness. RCO is a research format, not a Sharp replacement. See the companion hybrid K-quant repository if maximum quality under 1.8 GB is the goal.

Model summary

Item	code16	allmlp4	downq5
Base model	`openbmb/MiniCPM5-2B`	same	same
Architecture	LlamaForCausalLM, 42 layers	same	same
Parameters	2.52 B	same	same
Linear parameters	1.982 B	same	same
Attention	RCO3/4	168 x Q6_K	168 x Q6_K
MLP gate/up	RCO3/4	84 x RCO4	84 x RCO4
MLP down	RCO3/4	42 x RCO4	42 x Q5_K
Walls	Q6_K (2)	Q6_K token + Q8_0 output	Q6_K token + Q8_0 output
Norms	F32 (85)	F32 (85)	F32 (85)
Total tensors	381	381	381
RCO tensors	294	126	84

Evaluation

All numbers below use the same harness and the same held-out corpus:

WikiText-2 test, llama-perplexity -c 512 --chunks 16.
Held-out Python stdlib code (code2), -c 512 --chunks 32, 8176 scored tokens, KL against F16 donor logits. code2 was not used for calibration.
128 KB Sharp-style proxy: the first 131,041 bytes of the same held-out code corpus, 64 chunks.

Variant	Size	Wiki PPL	code2 mean KL	code2 p99	code2 top-1	128 KB p99	128 KB top-1
`code16`	1.431 GB	14.9480	0.451279	3.716013	78.824%	3.486164	78.641%
`allmlp4`	1.677 GB	14.3008	—	2.554672	83.137%	2.618498	83.090%
`downq5`	1.759 GB	14.0746	0.197163	1.952683	86.091%	1.952128	85.957%
Sharp Q4_K_XL	1.596 GB	14.1703	0.117031	1.236705	89.314%	1.167117	89.060%

downq5 is the best RCO-lineage model in this repository. It still trails Sharp Q4_K_XL on code, but it is better than Sharp Q4_K_XL on WikiText-2 PPL.

Quantization recipe

code16: RCO GPTQ/FastOBQ database, 128-weight RCO blocks, group size 128, per-layer 3/4-bit assignment from a 16x512 code search. Attention is RCO, walls are Q6_K.
allmlp4: same RCO database, but all 126 MLP tensors are RCO4; attention is exported as F16 and quantized to Q6_K with a 32x512 imatrix; token is Q6_K; output is Q8_0.
downq5: same as allmlp4, except down_proj is exported as F16 and quantized to Q5_K with the same imatrix. Gate/up remain RCO4.

The RCO block layout is 128 weights per block, BF16 scale d, BF16 zero-point z, and reconstructed weights y = d * (q - z). See docs/QUANTIZATION.md for the exact type files and commands.

Runtime

The RCO tensors require a llama.cpp v0.7.0 build with the patch from runtime/ applied. Stock v0.7.0 rejects the files; see runtime/README.md for the exact steps and the verified build commands.

./build-rocm/bin/llama-cli \
  -m minicpm5-2b-rco-downq5-q6attn-q8out-q6tok.gguf \
  -ngl 99 -c 512 -n 64 -t 8 --jinja \
  -p "Write a Python function that reverses a linked list."

downq5 contains 84 RCO4 tensors and 42 Q5_K tensors; allmlp4 contains 126 RCO4 tensors; code16 contains 66 RCO3 + 228 RCO4 tensors.

Speed on the patched v0.7.0 ROCm build (gfx1200, 8 threads, llama-bench -p 32 -n 8 -r 2): pp32 1211.53 ± 17.77 t/s, tg8 115.45 ± 15.96 t/s.

Files

minicpm5-2b-rco-code16-q6walls.gguf
minicpm5-2b-rco-allmlp4-q6attn-q8out-q6tok.gguf
minicpm5-2b-rco-downq5-q6attn-q8out-q6tok.gguf
metadata/assign_code16.json
metadata/assign_allmlp4.json
metadata/downq5_types.txt
metadata/layer_bitwidths_code16.csv
metadata/layer_bitwidths_allmlp4.csv
metadata/minicpm5_2b_imatrix.gguf
metadata/final_summary.json
docs/QUANTIZATION.md
docs/USAGE.md
docs/RAPOR.tr.md
runtime/llama.cpp-v0.7.0-rco-minicpm5.patch
runtime/README.md
SHA256SUMS

Limitations

Experimental format. RCO3/RCO4 are not upstream GGML types; use the patch, not a random llama.cpp build.
ROCm tuning. The GPU path is tuned for gfx1200 (RX 9060 XT, 16 GB). Other GPUs should work but may need different chunking/thread settings.
Lossy. These are 3/4-bit / Q5_K quantizations; use the BF16 base model when exact outputs are required.
Code2 is held out. Do not tune on code2. Use a separate calibration split for further experiments.
Not Sharp-class on code. Sharp Q4_K_XL remains better on code tail at similar size. The hybrid K-quant repository is the better quality choice if RCO compatibility is not required.

License

The base model is Apache-2.0. The quantization metadata and documentation in this repository are released under Apache-2.0. llama.cpp and the RCO patch are MIT-licensed.