Vontra/Qwen3.8-Flash-Next-MLX-oQ2-MTP

Qwen3.8 Flash Next, MLX oQ2 with native MTP

A sensitivity-aware mixed-precision MLX conversion of Qwen/Qwen3.8-Flash-Next, quantised from the official BF16 checkpoint with its native MTP draft block preserved.

Original model · Qwen overview · MLX-VLM · Qwen Community License 1.0

About this conversion

This release uses a 2-bit affine base and assigns more precision to sensitive modules. The final allocation includes 3, 4, 5, 6, and 8-bit modules. The native Qwen4Exp MTP block uses the same mixed-precision approach and remains part of the checkpoint.

Item	Value
Repository	Vontra/Qwen3.8-Flash-Next-MLX-oQ2-MTP
Base model	Qwen/Qwen3.8-Flash-Next
Source weights	Official BF16 checkpoint
Format	MLX safetensors
Quantisation	Sensitivity-aware oQ2 mixed precision
Base allocation	2-bit affine, group size 32
Module overrides	297 at 3-bit, 60 at 4-bit, 36 at 5-bit, 129 at 6-bit, 224 at 8-bit
Override group sizes	32, 64, or 128 according to module
Native MTP	Included, one Qwen4Exp draft block
Indexed tensors	3,747 total, including 76 MTP entries
Weight shards	15
Weight size	70.782 GB / 65.921 GiB
Configured context	262,144 tokens
Architecture	qwen4_exp vision-language sparse MoE

The upstream tokenizer, chat template, image and video processor configuration, generation configuration, licence, and native MTP configuration are included.

Use a runtime with explicit qwen4_exp and native-MTP support. A runtime that does not construct the Qwen4Exp MTP module may reject the 76 MTP entries during strict loading.

Download and use

python -m pip install --upgrade huggingface_hub

hf download Vontra/Qwen3.8-Flash-Next-MLX-oQ2-MTP \
  --local-dir ./Qwen3.8-Flash-Next-MLX-oQ2-MTP

Add the downloaded directory to a compatible oMLX model directory and refresh the model registry. Native MTP is optional. Keep it disabled by default for this release because the measured depth-three path was slightly slower than baseline.

Apple M3 Studio performance

The validation used three measured 512-token runs per mode. The table reports median decode throughput.

Runtime mode	Runs	Output per run	Median generation speed
Native MTP disabled	3	512 tokens	26.1398 tokens/s
Native MTP enabled, depth 3	3	512 tokens	25.9290 tokens/s

Native MTP changed median throughput by -0.81% in this test. The telemetry sample accepted 33 of 57 draft proposals, an acceptance rate of 57.89%. Each of the three paired MTP-off and MTP-on 512-token runs produced matching output hashes.

Instruction following, factual recall, arithmetic, and coherent long-generation gates passed in both modes. MTP worked correctly, but it did not improve throughput on this checkpoint and test. The recommended default is therefore MTP disabled.

The benchmark covers text generation. Results vary with prompt length, context growth, cache state, runtime version, memory pressure, and thermal conditions.

Architecture

Qwen3.8 Flash Next combines Gated DeltaNet, Qwen Sparse Attention, sparse mixture-of-experts layers, widened gated residual streams, hashed bigram and trigram embeddings, and a native next-token-prediction block for speculative decoding.

Architecture detail	Upstream value
Language-model parameters	125B total / 6B active
N-gram embedding	51B parameters, 20,000,000 entries
Native MTP	4B parameters, one draft layer
Hidden size	2,560
Layers	48
Routed / active experts	512 / 10, plus 1 shared expert
Native context	262,144 tokens, extensible upstream to 1,000,000

For upstream evaluations, intended use, safety guidance, and the full architecture discussion, see the original model card.

Conversion and validation

• The converter read the official BF16 checkpoint directly.
• The recipe used a 2-bit affine base with 746 mixed-precision module overrides.
• Full checkpoint validation covered all 3,747 indexed tensors, all 15 shards, and all 76 native-MTP tensor entries.
• Three deterministic 512-token runs passed in each MTP mode with exact paired output parity.
• Instruction, factual, arithmetic, and coherence checks passed in both modes.
• Native-MTP telemetry captured the draft and acceptance counters.

This is a community conversion, not an official Qwen release.

Limitations

• oQ2 is an extreme-compression format. Evaluate instruction following, reasoning, factual accuracy, and visual understanding on the intended workload before deployment.
• Native MTP was 0.81% slower in the measured test. Acceptance alone does not guarantee a throughput gain.
• The configured 262,144-token context does not guarantee that every Apple-silicon system can run that length within available unified memory.
• This is an MLX checkpoint for Apple silicon. It is not a GGUF, CUDA, TensorRT-LLM, or vLLM checkpoint.
• Upstream model limitations and safety considerations still apply.

Licence and attribution

The upstream model is released under the Qwen Community License 1.0. The required licence text is included in this repository and should be reviewed before use or redistribution.

Model design, training, evaluations, and upstream documentation belong to Qwen and the original contributors. The MLX conversion, native-MTP preservation, Apple-silicon validation, and packaging are provided by Vontra.