Vontra/Qwen3.8-Flash-Next-MLX-4bit

Qwen3.8 Flash Next — MLX 4-bit

A native Apple-silicon conversion of Qwen/Qwen3.8-Flash-Next, quantised directly from the official BF16 checkpoint.

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

About this conversion

This repository contains a 4-bit affine MLX conversion of Qwen3.8 Flash Next. It was produced directly from Qwen's BF16 weights using group size 32. The smaller group is intentional: it also covers the model's 160-wide hashed n-gram embedding tables instead of leaving them in BF16.

Item	Value
Base model	Qwen/Qwen3.8-Flash-Next
Format	MLX safetensors
Quantisation	4-bit affine, group size 32
Conversion stack	mlx-vlm 0.6.3, mlx 0.32.0
Weight shards	22
Weight size	111.58 GB (103.91 GiB)
Configured context	262,144 tokens
Architecture	qwen4_exp vision-language sparse MoE

The upstream tokenizer, chat template, vision processor, and generation configuration are preserved. The optional upstream MTP head is not included in this checkpoint.

Qwen3.8 Flash Next uses the new qwen4_exp architecture. Use an oMLX or MLX-VLM build that explicitly lists qwen4_exp support. Older MLX-VLM releases cannot load this checkpoint.

Do not attach a Qwen3.8 27B MTP drafter to this model. The hidden sizes differ and the drafter is incompatible with Flash Next.

Quick start

hf download Vontra/Qwen3.8-Flash-Next-MLX-4bit \
  --local-dir Qwen3.8-Flash-Next-MLX-4bit

With a compatible MLX-VLM runtime:

python -m mlx_vlm.generate \
  --model Qwen3.8-Flash-Next-MLX-4bit \
  --prompt "Explain sparse mixture-of-experts routing." \
  --max-tokens 512

Measured performance

Validated on an Apple M3 Studio with text-only generation after model load:

Test path	Result
oMLX server, warmed 543–566-token responses	24.1–24.2 tokens/s
oMLX server, warmed shorter responses	24.6–26.1 tokens/s
Standalone MLX exact-copy smoke test	31.0 tokens/s

The standalone result is a short smoke test; the longer oMLX figures better represent sustained chat generation. Results vary with prompt length, cache state, sampling settings, runtime version, and memory pressure.

Architecture

Qwen3.8 Flash Next is an experimental vision-language architecture combining Gated DeltaNet, Qwen Sparse Attention, sparse mixture-of-experts layers, widened gated residual streams, and hashed bigram/trigram embeddings.

Architecture detail	Upstream value
Language-model parameters	125B total / 6B active
N-gram embedding	51B parameters
Layers	48
Routed / active experts	512 / 10, plus 1 shared
Attention heads / KV heads	24 / 2
Hidden size	2,560
Native configured context	262,144 tokens

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

Conversion and validation

• Source: official BF16 checkpoint.
• All 3,671 converted tensors and 22 indexed shards were checked locally.
• The release payload was scanned for credentials, personal contact details, private paths, private network information, logs, caches, and private organisation data.
• Deterministic standalone and warmed oMLX server generation tests passed on Apple silicon.
• Quantisation can reduce output quality relative to BF16. Test the model on representative workloads before production use.

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

License and attribution

The upstream model is released under the Qwen Community License 1.0. The required licence text is included in this repository.

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