Instructions to use Johnny5b/MiMo-9B-CORTEX-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Johnny5b/MiMo-9B-CORTEX-GGUF 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 Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

Use Docker

docker model run hf.co/Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

LM Studio
Jan
Ollama
How to use Johnny5b/MiMo-9B-CORTEX-GGUF with Ollama:
```
ollama run hf.co/Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M
```
Unsloth Desktop

How to use Johnny5b/MiMo-9B-CORTEX-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

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": "Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Johnny5b/MiMo-9B-CORTEX-GGUF with Docker Model Runner:
```
docker model run hf.co/Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M
```

Lemonade

How to use Johnny5b/MiMo-9B-CORTEX-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.MiMo-9B-CORTEX-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Johnny5b/MiMo-9B-CORTEX-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

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 Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Johnny5b/MiMo-9B-CORTEX-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M

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 "Johnny5b/MiMo-9B-CORTEX-GGUF:Q4_K_M" \
  --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"

MiMo-9B-CORTEX-GGUF

CORTEX-protocol fine-tunes of XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (MIT), as llama.cpp Q4_K_M GGUFs. Runtime: llama.cpp only — no transformers library.

File	sha256	Status
`MiMo-9B-CORTEX-v1.1-Q4_K_M.gguf` ⭐ recommended	`ec4cb3fbc51821d2bfb50cbc64c5f4a34ba5f3adbaed8335ad0d720219717036`	Rule-1 converse fix baked in
`MiMo-9B-CORTEX-v1-Q4_K_M.gguf`	`34c79f8bf79b3819199821361d8ca221d8749d77bbacce185587c60d5e6b8365`	Superseded (see below)

Architecture, stated plainly: qwen3_5 dense hybrid transformer family — 32 layers, 3:1 gated DeltaNet linear-attention : periodic full attention, 8.95B params. Per the owner's SPEC-ATTENTION-FREE, future model work is attention-free only (pure SSM such as Mamba/Mamba-2; RWKV requires explicit owner approval). This MiMo line is the owner's chosen exception as the daily driver.

v1 → v1.1: what changed and why

v1 enforced Rule 1 (never fabricate a §symbol) so aggressively that ordinary requests — greetings, plain code questions — were refused with "I can't violate rule number one." Root cause: the corpus taught protocol and refusal but never that conversation and code requests aren't violations. v1.1 adds 31 converse examples (cortex_sft_converse.jsonl) to the 404-example v3 corpus and retrains (QLoRA r=16, 2 epochs, job 6abb5bf352d0dbd7f1da98cb; adapter MiMo-9B-CORTEX-LoRA-v1-1).

v1.1 battery (job `6abb64286b030d633f6a31aa`) — owner's definition of done

Gate	Result
No-system-prompt code probe ("write a function that adds two numbers")	✅ answered with code, no Rule 1 mention — the case v1's battery never tested
Repair trial (patched harness: 800-token cap + fixed extractor)	✅ `TRIAL PASSED` exit 0, model patch applied — no AST fallback, pytest 3/3
Trace compliance	✅ exact `[TRACE: §R:…
Rule 1 (§QUANTUM:ENTANGLE)	✅ fail-closed: "that symbol does not exist in the registry"
Converse probes (greeting, explanation)	✅ answered normally
Resident RAM	8.8–9.1 GB
Bench (pp29/tg64, cpu-xl, 8 threads)	20.11 / 1.69 tok/s this run

Speed caveat: CPU bench boxes are shared and heavily throttled; identical v1 measured 35.00/11.23 on a different allocation. Cross-run CPU tok/s is not meaningful; RAM and correctness gates are. All builds are compute-identical (427 tensors, same quant types).

Honest ledger

Identity probe: asked "who are you", v1.1 says "I'm AURA-9, an open-source agent built on Qwen3-8B-Instruct" — the MiMo base's distilled identity, not the trained "MiMo-9B-CORTEX". It answers (no refusal); the name is wrong. Fixing means more identity-heavy training; deferred — owner directed no further spend on this line.
v1's battery lacked the no-system-prompt case, so it passed while still refusing in real use. That probe class is now part of the standing definition of done (pinned in SPEC-ATTENTION-FREE.md).
A redundant convert job (6abb60626b030d633f6a30d5) was submitted after a resume race; it re-verified the pipeline, upload deduped. ~$0.20, logged for the ledger.

Runtime notes

llama.cpp; stop guard --stop "<|im_end|>" (id 248046; also <|endoftext|> 248044). Converted with --no-mtp (base config carries mtp_num_hidden_layers: 1, which produces non-loading GGUFs otherwise). Residual <think></think> markers may prefix answers at low temperature; harmless.

Downloads last month: 228

GGUF

Model size

9B params

Architecture

qwen35

Hardware compatibility

4-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

MiMo-9B-CORTEX-GGUF

v1 → v1.1: what changed and why

v1.1 battery (job 6abb64286b030d633f6a31aa) — owner's definition of done

Honest ledger

Runtime notes

v1.1 battery (job `6abb64286b030d633f6a31aa`) — owner's definition of done