Instructions to use AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit 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("AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit")

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 AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit"

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": "AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

OpenClaw new

How to use AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit"

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 "AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit" \
  --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"

MLX LM

How to use AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit 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 "AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit"

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 AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit

Run Hermes

hermes

Atomic Chat

AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit

Experimental AXQ 2-bit MLX pack of Qwen/Qwen3.8-2.4T-A95B-FP8.

Not certified. Will not be certified in this revision. Layer-stack SSD expert paging is too slow for practical serving, so this pack is a hobby / curiosity artifact: a 2.4T-class Qwen MoE that can exist on a Mac only because experts are paged from disk. If that sounds fun, enjoy. If you need something you can actually work with, use a smaller certified AXQ pack (Qwen 3.6, Flash, Coder-Next, GPT-OSS). No AXQ 4-bit sibling will be published for this base.

This card is convert evidence, not a quality or speed claim. Quality vs BF16 / FP8 was not measured.

This is an AXQuant pack (qwen38-moe-v1), not mlx-optiq. Do not load the OptiQ Qwen 3.8 repos in AX Engine.

Full convert notes: docs/qwen38-axq-2bit.md.

Why it is slow

The full table is ~1.13 TiB. No shipping Mac can resident-load it (512 GB unified memory is still too small). AX Engine pages one fused expert layer at a time (ax_expert_stream.json, required=true). Every token waits on SSD I/O for routed experts. That is why this revision is not a product path.

Estimated streamed peak: ~56 GiB resident trunk + ~12 GiB one layer + KV (about 100 GiB with headroom). You still need AX_ENGINE_2BIT_EXPERIMENTAL=1.

Do not mlx_lm.load this pack as a fully resident model.

Recipe

Affine, group size 32 (qwen38-experimental-2bit-v0.1.yaml):

Role	Bits
Expert, attention, shared MLP	2
Embedding, router (`mlp.gate`)	8
Norms, LM head	16 (BF16)
MTP	16, byte-preserved into `mtp.safetensors`

Source: official FP8 revision d2dc35658bcf77e66643428cb52e774cc3b5bd29 (128×128 weight_scale_inv, unfused per-expert tensors). Convert uses the AXQuant stream backend because mlx_lm.load cannot ingest that snapshot.

Measured artifact

From axquant_manifest.json after weight verification (2026-08-13):

Quantity	Value
Plan effective BPW	3.157
Measured total BPW	4.074
Measured main BPW	4.030
Logical parameters	2,446,182,725,504
Weight files	1,245,853,341,088 bytes (1.13 TiB)
MTP sidecar	26,989,794,744 bytes
Product class	`2bit-experimental`

Stream contract: 512 experts, 10 per token, 828 streamed tensors, layer-stack mode. Converted model_type is qwen3_5_moe (92 hybrid-attention layers, hidden 8192, MoE intermediate 2048, one MTP layer).

Status

Item	Status
Convert + `ax_expert_stream.json`	Done on `df-macstudio-m2`
Hub weights	Uploaded
Quality vs BF16 / FP8	Not measured
AX Engine cert	Will not certify this revision (too slow to be practical)
MTP acceleration	Not claimed

Hobby use only. Not a supported product pack.

Downloads last month: -

Safetensors

Model size

305B params

Tensor type

F32

U32

MLX

Hardware compatibility

2-bit

Model tree for AutomatosX/AX-Qwen3.8-2.4T-A95B-MLX-AXQ-2bit

Base model

Qwen/Qwen3.8-2.4T-A95B

Quantized

Qwen/Qwen3.8-2.4T-A95B-FP8

Quantized

(2)

this model