Instructions to use LyJonathan/MiniMax-M3.1-Flash-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 LyJonathan/MiniMax-M3.1-Flash-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 LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M
# Run inference directly in the terminal:
llama cli -hf LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M
# Run inference directly in the terminal:
llama cli -hf LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_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 LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M
# Run inference directly in the terminal:
./llama-cli -hf LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_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 LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M

Use Docker

docker model run hf.co/LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M

Llama new
LM Studio
Jan
Ollama
How to use LyJonathan/MiniMax-M3.1-Flash-GGUF with Ollama:
```
ollama run hf.co/LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M
```
Unsloth Desktop

How to use LyJonathan/MiniMax-M3.1-Flash-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_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": "LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use LyJonathan/MiniMax-M3.1-Flash-GGUF with Docker Model Runner:
```
docker model run hf.co/LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M
```

Lemonade

How to use LyJonathan/MiniMax-M3.1-Flash-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M

Run and chat with the model

lemonade run user.MiniMax-M3.1-Flash-GGUF-IQ1_M

List all available models

lemonade list

Hermes Agent

How to use LyJonathan/MiniMax-M3.1-Flash-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 LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_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 LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use LyJonathan/MiniMax-M3.1-Flash-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_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 "LyJonathan/MiniMax-M3.1-Flash-GGUF:IQ1_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"

MiniMax-M3.1-Flash GGUF

EXPERIMENTAL. This project is no longer under active development as of 2026-10-26.

Quantized GGUF copies of MiniMax-M3.1-Flash for Sparsa. Most of these files do not load in stock llama.cpp, LM Studio or Ollama; use Sparsa, latest release v0.1.3.

Tiers

Each tier is a folder of 13 parts; point Sparsa at the folder. Quality is measured against an FP32 reference of the original weights, 6 short prompts, 679 positions: top-1 is how often the tier's top next token matches the reference (higher is better), KL64 is divergence over the reference's top 64 tokens (lower is better).

Tier	Size	Top-1	KL64	Use
IQ1_S	94.8 GB	44.3 %	1.926	Smallest tier; do not use, its output is incoherent (below)
IQ1_M	106.6 GB	47.1 %	1.673	Smallest usable tier; the 1-bit tiers are noticeably degraded
IQ2_KT	123.8 GB	56.3 %	1.292	Best measured 2-bit per byte; Sparsa only
IQ2_XXS	124.5 GB	52.0 %	1.804	2-bit that also loads in patched llama.cpp
IQ2_XS	133.1 GB	quality not measured	-	2-bit for patched llama.cpp; no Sparsa measurement (note below)
IQ2_XS-D3	146.1 GB	59.4 %	1.211	Default pick: best measured 2-bit, runs on a 128 GB PC
IQ2_KL	152.7 GB	59.8 %	1.184	2-bit, slightly better than D3; Sparsa only
IQ3_K	191.2 GB	69.8 %	0.847	3-bit step up from the 2-bit tiers; Sparsa only
IQ4_KS	233.6 GB	74.4 %	0.760	Closest IQ tier to NVFP4; Sparsa only
NVFP4	246.1 GB	77.3 %	0.670	Best measured quality, a lossless copy of the checkpoint's 4-bit experts; Sparsa only

IQ2_XS is only measured in patched llama.cpp on a harder eight-prompt set (top-1 46.8 %, KL 1.92; IQ2_XS-D3 on that set: 53.8 %, 1.58), so the number is not comparable with the table. Sparsa only means upstream llama.cpp cannot read those expert types. The experts were trained as 4-bit (NVFP4), so the 1-2-bit tiers are degraded, and one hard code prompt pulls every mean down.

Pick the largest tier that fits in RAM + VRAM with ~10 GB to spare. Sparsa streams the rest from the SSD, which works but is slower.

IQ1_S is broken in practice: in our test it loops ("and and and...") instead of writing text. Use IQ1_M or larger.

Running with Sparsa

git clone https://github.com/LyJonathan/Sparsa.git && cd Sparsa
# unzip sparsa-v0.1.3-windows-x64-cuda.zip (or the linux tarball) into bin/, then:
python -m pip install -r requirements.txt
python sparsa.py download iq2-xs-d3 --dir D:/models
python sparsa.py download tokenizer --dir D:/models/MiniMax-M3.1-Flash-IQ2_XS-D3/tokenizer
python sparsa.py serve --model D:/models/MiniMax-M3.1-Flash-IQ2_XS-D3

That serves OpenAI- and Anthropic-compatible APIs on http://127.0.0.1:8080.

Vision. mmproj-MiniMax-M3.1-Flash-F16.gguf (1.73 GB, in this repo) adds image input for every tier: python sparsa.py download mmproj --dir <tier folder>. The encoder runs on the CPU in fp32 and takes 64-80 s per 640x480 image; images only, no video.

Measured speed (IQ2_XS-D3, 2x RTX 5070 12 GB, Ryzen 9 5900X, 128 GB DDR4, one NVMe): about 4 tok/s writing, 13 tok/s reading a 400-token prompt, 44 tok/s on a 32K prompt. Speed depends heavily on the SSD.

Known limitations

IQ1_S produces incoherent output; the warning above stands.
Rare stall. About 5 % of runs stall in the CPU-expert prompt path. The cause is not known yet. Because of it, PR #34 (+47 % prompt speed when prompt experts run on CPU) is not merged.
Vision is CPU-only and slow, and takes images only.
Context. 8K-32K works; above 32K expect it to be slow or to fail.
Maintenance. No further development as of 2026-10-26.

How they are made

Routed experts are quantized at the listed type with an imatrix from 23,738 calibration tokens; the small tiers put ffn_down_exps one type higher.
Attention, indexer, dense layers and shared experts stay at Q8_0 (Q6_K in IQ1_S/IQ1_M); output is Q6_K.
NVFP4 is a lossless copy of the checkpoint's 4-bit experts.
llama.cpp (patched) needs the small MiniMax-M3 patch that ships with Sparsa.
Checksums: SHA256SUMS (sha256sum -c SHA256SUMS).

License

The weights are under the MiniMax Community License (LICENSE): commercial use requires displaying "Built with MiniMax M3.1-Flash" and a notice to api@minimax.io, and its prohibited uses apply.

Credits

MiniMax (model, chat template, vision encoder); Strata by Niko1221 (engine base); ggml / llama.cpp (GGUF format, quantization, kernels); ik_llama.cpp by Iwan Kawrakow (IQ2_KT/IQ2_KL/IQ3_K/IQ4_KS formats); Project Maya by Peasant Smith (GPTQ error-feedback rounding, tier ideas); agentionai's llama.cpp (trellis types); Unsloth's dynamic quants and ubergarm's ik_llama.cpp recipes (small-tier mix, no code copied); stb_image. LICENSE and THIRD_PARTY_NOTICES.md in the Sparsa repository carry the full notices.

Downloads last month: 13,885

GGUF

Model size

426B params

Architecture

minimax-m3

Hardware compatibility

1-bit

2-bit

4-bit

Inference Providers NEW

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

Model tree for LyJonathan/MiniMax-M3.1-Flash-GGUF

Base model

MiniMaxAI/MiniMax-M3.1-Flash

Quantized

(32)

this model

Quantizations

1 model