Instructions to use Joonix71/Apodex-1.1-mini-MTPLX-Balance with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Joonix71/Apodex-1.1-mini-MTPLX-Balance 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("Joonix71/Apodex-1.1-mini-MTPLX-Balance")

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 Joonix71/Apodex-1.1-mini-MTPLX-Balance with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Joonix71/Apodex-1.1-mini-MTPLX-Balance"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Joonix71/Apodex-1.1-mini-MTPLX-Balance"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

MLX LM

How to use Joonix71/Apodex-1.1-mini-MTPLX-Balance with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "Joonix71/Apodex-1.1-mini-MTPLX-Balance"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "Joonix71/Apodex-1.1-mini-MTPLX-Balance"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "Joonix71/Apodex-1.1-mini-MTPLX-Balance",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Hermes Agent

How to use Joonix71/Apodex-1.1-mini-MTPLX-Balance 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 "Joonix71/Apodex-1.1-mini-MTPLX-Balance"

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 Joonix71/Apodex-1.1-mini-MTPLX-Balance

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Joonix71/Apodex-1.1-mini-MTPLX-Balance with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "Joonix71/Apodex-1.1-mini-MTPLX-Balance"

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 "Joonix71/Apodex-1.1-mini-MTPLX-Balance" \
  --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"

Apodex-1.1-mini MTPLX Balance

An MTPLX pack of apodex/Apodex-1.1-mini for native multi-token-prediction (MTP) speculative decoding on Apple Silicon.

Apodex-1.1-mini is an agentic fine-tune of Qwen3.5-35B-A3B by Apodex AI. The source release ships an MTP layer. Existing MLX conversions drop that layer, so they can only decode autoregressively. This pack keeps the MTP layer in BF16, so MTPLX can draft with it.

Measurements

M5 Pro, 64 GB, 2026-09-29. The Forge verification ran on MTPLX 2.12.0. The server workload ran on 2.12.0 plus the open pull requests to MTPLX with prefill and session-cache speed-ups (#549, #550, #555-#559, #570); those mainly shorten prefill and barely move decode.

Forge verification (mtplx tune, suite long-code-uncapped, 2,048 tokens, thinking off):

Mode	tok/s	vs AR	Acceptance per position
AR	64.5	1.00
D1	91.5	1.42	0.941
D2	111.0	1.72	0.917 / 0.777
D3	98.5	1.53	0.915 / 0.776 / 0.671

Server workload at depth 2, turbo profile, temperature 0.6: decode 93.9 tok/s over 256 tokens, 111 tok/s on short replies. At temperature 1.0 top-p 0.95 (the model card's recommendation) the speed was the same.

Speculative decoding with MTPLX's exact acceptance keeps the target model's output distribution; the head only changes speed.

Build

Built with mtplx forge build from the BF16 source. Recipe: body_bits 6, body_group_size 64, 8-bit overrides for mlp.gate and mlp.shared_expert_gate, and mtp_policy keep_bf16.
The source stores its MTP experts as numbered tensors, which MTPLX 2.12.0 loads directly.
The vision tower is included as in the source release.

Use

hf download Joonix71/Apodex-1.1-mini-MTPLX-Balance \
  --local-dir ~/.mtplx/models/Apodex-1.1-mini-MTPLX-Balance
mtplx serve --model ~/.mtplx/models/Apodex-1.1-mini-MTPLX-Balance --mtp --port 8000

Served model id: mtplx-apodex-1.1-mini-balance. Recommended depth: 2.

Apodex recommends temperature 1.0, top-p 0.95 and repetition penalty 1.05 for agentic work, with the Qwen3.5 chat template (qwen3_coder tool calls, qwen3 reasoning).

Settings we measured with

These settings exist in stock MTPLX 2.12.0 and match the server setup of the measurements above.

MTPLX_PREFILL_CHUNK_SIZE_DENSE=4096 MTPLX_PREFILL_CHUNK_SIZE_REPAGE=4096 \
  mtplx serve --model <pack dir> --mtp --profile turbo --depth 2 --port 8000

Depth 2 is also the pack's recommended depth, so --depth can be left out.
For long agent sessions with thinking on, MTPLX_THINKING_BUDGET caps runaway thinking blocks. We run 6144; that is a safeguard, not a measured speed-up.

License

Apache 2.0, as the source model. All credit for the model goes to Apodex AI and the Qwen team; this repository only repackages it for MTPLX.

Downloads last month: 171

Safetensors

Model size

35B params

Tensor type

U32

BF16

MLX

Hardware compatibility

6-bit

Model tree for Joonix71/Apodex-1.1-mini-MTPLX-Balance

Base model

Qwen/Qwen3.5-35B-A3B-Base

Finetuned

Qwen/Qwen3.5-35B-A3B

Finetuned

apodex/Apodex-1.1-mini

Quantized

(28)

this model