Instructions to use kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX"

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX"

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 kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX"

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 "kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

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

kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX

Mixed-precision (UVMAX) quantization of Qwen/Qwen3.8-2.4T-A95B, converted with mlx-lm from the bf16 release.

What is UVMAX?

UVMAX assigns bit widths per tensor class instead of quantizing uniformly.

tensor class	precision	parameters	size	share
Expert FFN gate/up	2-bit, group 128	1,580B	414 GiB	55.2%
Expert FFN down	3-bit, group 128	790B	299 GiB	39.9%
Attention + gated delta net	6-bit, group 64	40B	30.2 GiB	4.0%
Shared experts	8-bit, group 64	4.6B	4.6 GiB	0.6%
Embeddings, LM head	4-bit, group 64	4.1B	2.1 GiB	0.3%
Routers	8-bit, group 64	0.4B	0.4 GiB	<0.1%
Norms, convolutions, gates	bf16 / fp32	0.01B	—	<0.1%
total	2.66 bits/weight	2.42T	750 GiB

Quality

Teacher-forced against the bf16 release on identical tokens, 48 windows of 1025 tokens. KLD is KL(bf16 ‖ UVMAX) over the full output distribution and is corpus-specific.

corpus	ppl bf16	ppl UVMAX	ratio	mean KLD	median KLD	top-1 agreement
Linux kernel C	1.300	1.406	1.08×	0.099	0.0008	94.9%
XNU kernel C	2.122	2.272	1.07×	0.106	0.0061	91.7%
JavaScriptCore C++	2.059	2.184	1.06×	0.102	0.0031	92.1%
English prose	2.684	3.188	1.19×	0.225	0.0528	84.8%
all	1.976	2.172	1.10×	0.133	0.0062	90.9%

Throughput

2× M3 Ultra (512 GB), tensor parallelism over Thunderbolt RDMA, batch size 1, peak memory per node.

prompt tokens	prefill tok/s	generation tok/s	peak memory
4k	176	14.2	385 GiB
16k	171	13.7	385 GiB
32k	162	13.1	386 GiB
64k	142	12.3	391 GiB
128k	113	10.7	402 GiB

Use with mlx

Requires mlx-lm 0.31.3 or later.

pip install mlx-lm

The model does not fit on one 512 GB machine. Serve it across two with an MLX distributed hostfile:

mlx.launch --backend jaccl --hostfile hosts.json --env MLX_METAL_FAST_SYNCH=1 -- \
  mlx_lm.server --model kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX

Sampling follows the base model: temperature 1.0, top-p 0.95, top-k 20.

Downloads last month: 1,044

Safetensors

Model size

221B params

Tensor type

BF16

U32

MLX

Hardware compatibility

2-bit

Model tree for kernelpool/Qwen3.8-2.4T-A95B-3bit-UVMAX

Base model

Qwen/Qwen3.8-2.4T-A95B

Quantized

(26)

this model