Qwopus3.6-35B-A3B-v1-oQ8-MTP

Unofficial oQ8 MLX/oMLX quantized conversion of Jackrong/Qwopus3.6-35B-A3B-v1. This model was quantized using oQ / oMLX mixed-precision quantization with oMLX v0.3.9. This conversion preserves the model's native MTP support and is intended for oMLX / MLX inference on Apple Silicon.

This is not an official Qwen, Alibaba, or QwOpus release.
It is an unofficial community quantized conversion for local inference.

Redownload notice

If you downloaded an earlier revision of this repository, please remove the old local cached copy and re-download the full repository. This conversion includes oQ8 weights and native MTP-related files/configuration. Mixing old config, tokenizer, or model shard files with a newer revision may cause loading errors, MTP not working correctly, or unstable generation.

Base model

• Base model: Jackrong/Qwopus3.6-35B-A3B-v1
• Base architecture: Qwen3.6-35B-A3B / qwen3_5_moe
• License: Apache-2.0, inherited from the base model
• Base model notes:
  • Sparse MoE model
  • 35B total parameters / approximately 3B active parameters per token
  • Native long-context support

Quantization details

• Model type: qwen3_5_moe
• Quantization method: oQ mixed-precision quantization
• Bits: 8
• Group size: 64
• Format: MLX safetensors
• Runtime used for conversion: oMLX v0.3.9
• Native MTP: preserved / intended to be used
• Additional fine-tuning: none

Runtime notes

Tested primarily with:

• Runtime: oMLX v0.3.9
• Hardware: Apple Silicon, tested on Mac Studio M4 Max 128GB
• Context tested: up to 262144 ctx
• MTP: ON
• DFlash: OFF Recommended initial runtime flags based on local testing:

MTP: ON
DFlash: OFF
SpecPrefill: OFF initially
In-memory DFlash cache: OFF
SSD DFlash cache: OFF

DFlash is not recommended for this conversion based on local testing, where it caused repetition / looping on agentic prompts.

Local performance notes

Local observations on Mac Studio M4 Max 128GB with oMLX v0.3.9:

Setting	Observed decode speed
MTP OFF	~58-59 tok/s
MTP ON	~60-70 tok/s
MTP ON average	~64 tok/s
MTP ON low-context peak	~70 tok/s

These are local observations, not standardized benchmarks. Actual speed depends on prompt length, context length, oMLX version, generation settings, temperature, cache behavior, and Apple Silicon memory pressure.

Format notes

This is not GGUF, AWQ, GPTQ, or bitsandbytes format.

This repository contains an MLX safetensors conversion intended for oMLX / MLX-compatible runtimes.

Known notes

* Native MTP showed a measurable decode-speed improvement in local testing.
* DFlash caused repetition / looping in local testing and is not recommended for this model.
* If generation becomes unstable, disable speculative features other than native MTP first.
* If loading fails after a repo update, delete the old local cache and re-download the full repository.
* This conversion is mainly tested on local agentic coding / advisor prompts, not on a full public benchmark suite.

License

This quantized conversion follows the license of the base model: Apache-2.0.

Apache-2.0 allows redistribution and modification, provided that the license and applicable notices are preserved and modifications are stated. This repository is an unofficial quantized conversion and does not claim affiliation with the base model author, Qwen, Alibaba, or any related organization.

Attribution

* Base model: Jackrong/Qwopus3.6-35B-A3B-v1
* Quantization / conversion: orangeseasun
* Quantization tool: oMLX

Changes from base model

Compared with the base model, this repository:

* converts the original model weights to oQ8 MLX safetensors
* preserves native MTP-related files/configuration
* does not add additional fine-tuning
* is packaged for local oMLX / Apple Silicon usage