Qwen3.8 Flash Next — MLX oQ4
A mixed-precision MLX 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 an oQ4 mixed-precision MLX conversion produced directly from Qwen's BF16 weights. Layer sensitivity was measured with a validated 4-bit proxy; final tensor quantisation reread the official BF16 checkpoint. Group size 32 supports the model's 160-wide hashed n-gram embedding tables.
| Item | Value |
|---|---|
| Base model | Qwen/Qwen3.8-Flash-Next |
| Format | MLX safetensors |
| Quantisation | oQ4 mixed precision; 4-bit affine base with protected layers at 5/8-bit |
| Base group size | 32 |
| Weight shards | 22 |
| Weight size | 111.69 GB (104.02 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.
[!IMPORTANT] Qwen3.8 Flash Next uses the new
qwen4_exparchitecture. Use an oMLX or MLX-VLM build that explicitly listsqwen4_expsupport. Older MLX-VLM releases cannot load this checkpoint.
[!CAUTION] 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 TensorFold/Qwen3.8-Flash-Next-MLX-oQ4 \
--local-dir Qwen3.8-Flash-Next-MLX-oQ4
With a compatible MLX-VLM runtime:
python -m mlx_vlm.generate \
--model Qwen3.8-Flash-Next-MLX-oQ4 \
--prompt "Explain sparse mixture-of-experts routing." \
--max-tokens 512
Measured performance
Validated on an Apple M3 Studio with deterministic text-only generation after model load:
| Test path | Result |
|---|---|
| Standalone MLX exact-copy generation | 30.4 tokens/s |
| Standalone MLX, 139-token explanatory response | 27.9 tokens/s |
| oMLX server, warmed 156-token response | 27.9 tokens/s |
| oMLX server, warmed 1,044-token response | 27.0 tokens/s |
Prompt processing measured 27.2–32.2 tokens/s in the standalone short tests. The first oMLX server request reported 6.72 seconds to load the model; that one-off load time is separate from generation speed. The longer warmed server run is the most representative sustained result. Results vary with prompt length, cache state, sampling settings, runtime version, and memory pressure.
Architecture
Qwen3.8 Flash Next combines 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.
- Sensitivity-guided mixed-precision allocation: 4-bit base with 228 protected modules at 5/8-bit.
- All 3,671 converted tensors and 22 indexed shards were checked locally.
- Deterministic exact-copy and explanatory-generation tests passed on Apple silicon.
- The release payload was scanned for credentials, personal contact details, private paths, private network information, logs, caches, and private organisation data.
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 TensorFold.
Choose for your Mac
64GB Macs · 128GB Macs · 256GB Macs
No measured memory tier is assigned here. The collections use published M3 Studio peaks with at least 25% nominal headroom; fit on other Macs is an estimate, and full context is not guaranteed. Start with short context and one request.
Runtime and evidence
The exact tested oMLX application version is not recorded here; a library version is not an app version. The original performance tables retain their benchmark conditions and speed figures; this documentation update adds no new test results.
An unresolved long-generation repetition report involves oQ4-MTP; this update does not establish a fix.
Quick start and demo prompt
hf download TensorFold/Qwen3.8-Flash-Next-MLX-oQ4 --local-dir ./models/Qwen3.8-Flash-Next-MLX-oQ4
Add the downloaded folder to oMLX model directories, refresh the list, and follow this card's architecture and MTP compatibility requirements before loading.
Try this in a new chat with a 128-token output limit:
Explain why the sky looks blue in three short sentences.
This is a demo prompt to try, not a recorded successful run; a captured demonstration for this documentation update is not yet available.