Instructions to use AmixDigital/BigBang-v1-mtp-oQ2.7e with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AmixDigital/BigBang-v1-mtp-oQ2.7e with MLX:

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

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

# Load the model
model, processor = load("AmixDigital/BigBang-v1-mtp-oQ2.7e")
config = load_config("AmixDigital/BigBang-v1-mtp-oQ2.7e")

# Prepare input
image = ["http://images.cocodataset.org/val2017/000000039769.jpg"]
prompt = "Describe this image."

# Apply chat template
formatted_prompt = apply_chat_template(
    processor, config, prompt, num_images=1
)

# Generate output
output = generate(model, processor, formatted_prompt, image)
print(output)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use AmixDigital/BigBang-v1-mtp-oQ2.7e with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "AmixDigital/BigBang-v1-mtp-oQ2.7e"

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": "AmixDigital/BigBang-v1-mtp-oQ2.7e"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent

How to use AmixDigital/BigBang-v1-mtp-oQ2.7e 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 "AmixDigital/BigBang-v1-mtp-oQ2.7e"

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 AmixDigital/BigBang-v1-mtp-oQ2.7e

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use AmixDigital/BigBang-v1-mtp-oQ2.7e with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "AmixDigital/BigBang-v1-mtp-oQ2.7e"

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 "AmixDigital/BigBang-v1-mtp-oQ2.7e" \
  --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"

Community oQ quantization · Apple Silicon

BigBang-v1-mtp-oQ2.7e

One complete enhanced oQ2.7e MLX quantization of BigBang-v1 MTP for local Apple Silicon inference.

oQ2.7eenhanced oQ level

14.259 GiBcomplete repository size

4 MLX shardsone intact artifact

128 GB reportedfamily smoke test

01 / 15

About This Release

BigBang-v1-mtp-oQ2.7e is an independent, standalone MLX oQ quantization of endless-frontier/BigBang-v1. This repository contains exactly one complete oQ2.7e level, not a mixture of alternative tiers.

Selection: browse the BigBang-v1 MTP oQ Quantizations collection to compare the nine standalone levels.
Published artifact: 4 MLX Safetensors shards plus the exact model, tokenizer, prompt, and multimodal preprocessing assets required by this tier.
Maintainer transformation: oQ quantization only; no fine-tuning, merge, behavior edit, DFlash, DSpark, draft weights, or acceleration companion was added.

Notice

Important Notices

This is a community quantization, not an official Endless Frontier release.
A basic family-level smoke test was reported on a MacBook Pro with M4 Max and 128 GB unified memory. The exact oQ tier, macOS version, and MLX runtime version were not recorded.
No independent benchmark, reasoning, development, tool-use, thinking, vision, or long-context evaluation is claimed for this exact oQ2.7e artifact.
Results reported for the source model are not results for this quantization.

02 / 15

Model Details

Format	MLX Safetensors
Modality	Image-text-to-text, inherited from BigBang-v1
Architecture	Qwen3.5 MoE conditional generation, inherited through BigBang-v1
Quantization	oQ2.7e; 2-bit default affine configuration, group size 64, with 402 enhanced per-tensor overrides
Weights	4 shards, 14.259 GiB total repository size
Context	262,144 tokens configured upstream; not independently long-context tested

This page intentionally exposes one selectable quantization level. The Hub can therefore calculate its MLX footprint without summing the other eight alternatives.

03 / 15

Provenance and Lineage

Published artifact: this complete MLX oQ2.7e quantization, prepared by AmixDigital.
Direct source: endless-frontier/BigBang-v1 at revision fe313c9.
Technical ancestor: Qwen/Qwen3.6-35B-A3B, credited by the direct source. It is a technical ancestor, not this release's visual reference.

No fine-tuning, merge, behavior edit, or acceleration model was added by AmixDigital.

04 / 15

Transformation Details

This release was quantized with the oMLX engine through its graphical interface. It uses the oQ2.7e enhanced plan: a 2-bit default affine quantization, group size 64, and 402 recorded per-tensor precision overrides.

The local calibration report identifies oqe_code_multilingual and records 523 entries, 481 applied entries, two missing entries, no mismatches, and one zero-count expert. This is an allocation record, not a quality evaluation.

The machine-readable report is intentionally excluded because it contains a personal local path. The oMLX version and exact build procedure were not recorded.

05 / 15

Requirements and Compatibility

Reported family test environment: MacBook Pro with M4 Max and 128 GB unified memory; basic loading and operation were reported, but the exact tier and software versions were not recorded.
Platform: macOS on Apple Silicon only.
Runtime: an MLX-compatible local runtime supporting the inherited architecture and image/video preprocessing assets.
Memory: minimum unified memory is not independently measured. Disk size is not a safe proxy for memory requirements.
Not claimed: Windows, Linux, CUDA, GGUF, vLLM, or server-GPU compatibility.

06 / 15

Usage

Use this repository as one intact MLX model directory.
Keep its index, all shards, tokenizer, chat template, preprocessor, and video preprocessor configuration files together.
Load it in a compatible local MLX workflow on Apple Silicon.
Begin with short prompts and modest context, then observe memory pressure before increasing the workload.

Prompting, Thinking and Tool Use

The included chat template contains image, video, thinking, and tool-message formatting branches. No end-to-end thinking or tool-use result is claimed for this exact quantization.

07 / 15

Evaluation and Performance

No independent evaluation has been completed for this exact oQ2.7e release. A family-level basic smoke test was reported on a MacBook Pro with M4 Max and 128 GB unified memory, but neither the exact oQ tier nor a metric was recorded. Throughput, latency, memory, reasoning, development, generation quality, vision, tool use, thinking, and long-context behavior remain unmeasured. Upstream scores must not be attributed to this artifact.

08 / 15

Intended Uses

Local experimentation and evaluation on compatible Apple Silicon Macs.
Image-text-to-text research and development with human review.
Comparing oQ storage levels under the user's own controlled workload.

09 / 15

Out-of-Scope Uses

Unsupervised high-stakes medical, legal, financial, employment, safety, or infrastructure decisions.
Autonomous destructive tool execution without sandboxing, review, and recovery controls.
Claims of compatibility not documented in this card.

10 / 15

Risks, Biases and Limitations

Model outputs may be inaccurate, biased, unsafe, fabricated, or unsuitable for a particular context.
Quantization can introduce changes beyond the limitations of the direct source.
The calibration record has two missing entries and one zero-count expert; it is not an assurance of quality.
Minimum memory and functional behavior remain unmeasured.

11 / 15

Recommendations and Mitigations

Evaluate the exact oQ level and intended prompts before use.
Keep human review for consequential outputs and generated code.
Sandbox tool access and protect credentials, private data, and the filesystem.
Monitor memory pressure and stop before macOS memory compression or swapping compromises the workload.

12 / 15

License and Responsible Use

The quantized artifacts in this repository are distributed under the OpenMDW-1.1 License, included as LICENSE.

The direct source is licensed under Apache-2.0. Its license text is retained as LICENSE-APACHE-2.0. Recipients and redistributors must comply with all applicable terms and notices.

13 / 15

Acknowledgements

Endless Frontier for BigBang-v1, the direct source.
AmixDigital for this MLX oQ distribution.

14 / 15

Model Card Contact

Contact AmixDigital on Hugging Face or open a repository discussion to report an error or propose a correction.

15 / 15

Repository Files

model-*.safetensors and model.safetensors.index.json: the complete oQ2.7e MLX shard set and its shard map.
config.json, generation_config.json, chat_template.jinja: model, quantization, generation, and prompt configuration.
tokenizer.json, tokenizer_config.json, merges.txt, vocab.json: tokenizer assets.
preprocessor_config.json and video_preprocessor_config.json: multimodal preprocessing assets.
LICENSE, LICENSE-APACHE-2.0, .gitattributes, and this README.md: distribution metadata.