What is the fastest way to run Qwen 3.8 Flash-Next on a Mac?

MTPLX, which shipped the first Apple Silicon backend for the family in 2.10.0 on 29 August 2026 and runs the model's own multi-token-prediction head as an exact speculative decoder. Through the MTPLX server on an M5 Max it decodes 61 tok/s plain and 63 to 76 tok/s with the MTP head depending on workload. On the pack cards' coding task at official Qwen 3.8 sampling (temperature 1.0, top-p 0.95, top-k 20), fans verified at max, single stream, Optimized Speed ran 73.5 tok/s against 43.8 plain autoregressive and Bare Speed ran 75.9 against 47.0.

How much RAM does Qwen 3.8 Flash-Next need on a Mac?

96 GB of unified memory or more. Optimized Speed is a 115.1 GB download including the 32 GB n-gram table, with about 83 GB of resident weights plus working set; Bare Speed is 106.3 GB with about 74 GB resident. The n-gram table streams from SSD by default, so the weights stay resident and the table does not have to.

Does MTP speculative decoding change Flash-Next's output?

No. Drafts from the MTP head are accepted with the probability-ratio rule and rejected drafts are resampled from the residual, so the output follows the model's own distribution at any temperature. The quantization is the only approximation.

Run Qwen 3.8 Flash-Next on a Mac with MTPLX: 125B MoE, 96 GB+

Two packs

Both packs keep the model's MTP head and run the same exact speculative path. They differ in quantization. Optimized Speed keeps the Qwen Sparse Attention projections at 8-bit, so the attention pathway that steers long contexts keeps its precision; it is the recommended build. Bare Speed puts every expert at flat 4-bit for the quickest Flash-Next speeds.

Pack	Quant	Download	Resident weights	Pick it for
Optimized Speed	Dynamic 4-bit, 64-weight groups; Qwen Sparse Attention projections at 8-bit	115.1 GB	~83 GB + working set	Recommended. Higher quality, slightly slower.
Bare Speed	Flat 4-bit, 64-weight groups, nothing promoted	106.3 GB	~74 GB + working set	Quickest Flash-Next speeds for chat and coding.

Both downloads include the 32 GB n-gram embedding table. It ships as a separate ngram-table.safetensors sidecar that MTPLX streams from SSD by default (resident is opt-in on very large machines), so the weights stay resident and the table does not have to. Both packs keep the vision tower. Context window: 262,144 tokens. The base model is Qwen/Qwen3.8-Flash-Next under the Qwen Community License; the upstream model card is preserved in each repo as README-upstream-qwen.md.

Measured speeds

From the two pack cards. M5 Max, fans verified at max, single stream, through the MTPLX server (mtplx serve), official Qwen 3.8 sampling (temperature 1.0, top-p 0.95, top-k 20), sampled output, through the Flash-Next backend that shipped in MTPLX 2.10.0 on 29 August 2026. Same coding task for both packs.

Run	Optimized Speed	Bare Speed
Coding task, MTP speculative decode (the default)	73.5 tok/s	75.9 tok/s
Same task, plain autoregressive	43.8 tok/s	47.0 tok/s
Speculative multiplier through the product serve path	1.7x	1.6x

The 2.10.0 release note gives the spread across workloads through the MTPLX server on an M5 Max: 61 tok/s plain decode, 63 to 76 tok/s with the MTP head. The 2.10.1 release note (30 August 2026) added block-sparse prefill for Flash-Next: peak memory on a 98k-token prompt fell from 91.4 to 83.0 GB, and a 262,144-token cold prompt completed at 87.4 GB peak, where 2.10.0 needed 119 GB or did not complete.

RAM and Macs

96 GB or more: both packs, with the n-gram table streaming from SSD. Resident weights about 83 GB (Optimized Speed) or about 74 GB (Bare Speed) plus working set. The 2.10.1 release confirmed that 96 GB Macs load Flash-Next.
Context window: 262,144 tokens. The memory governor (2.10.0) prints engine budget, weights, resolved context window and session bank in the serve banner; requests that cannot fit are refused up front with HTTP 507 (2.10.2, 1 September 2026).
Image input: PNG, JPEG and WebP since 2.10.1 (30 August 2026).
Smaller Macs: Qwen 3.8 27B on 32 GB or more; Qwen 3.5 9B or 4B on 16 GB. The app checks your Mac before recommending anything.

Install

Mac app: download the DMG, pick "Qwen 3.8 Flash-Next Optimized Speed" or "Qwen 3.8 Flash-Next Bare Speed". The app downloads the pack, sets up its engine, and measures your machine to pick the fastest decoding depth.

Command line:

brew install youssofal/mtplx/mtplx
mtplx serve --model Youssofal/Qwen3.8-Flash-Next-MTPLX-Optimized-Speed

For Bare Speed, pass --model Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed. Then point OpenCode, Pi, Claude Code, Cline, Cursor or anything that speaks the OpenAI or Anthropic API at http://127.0.0.1:8000. The served model ids are mtplx-flash-next-optimized-speed and mtplx-flash-next-bare-speed; mtplx connect claude-code and mtplx connect opencode print the exact client config. Setup pages for each client are in the docs. The serving contract ships inside mtplx_runtime.json; MTPLX reads it on load. The Turbo profile is the default for the Flash-Next packs.

How it is built

Optimized Speed:

MoE experts and dense matrices at 4-bit with 64-weight groups; the Qwen Sparse Attention projections promoted to 8-bit, the quality edge over Bare Speed.
The GDN convolution and recurrent-state parameters, every norm, the QSA indexer, and the MTP head stay 16-bit.
The n-gram embedding table ships as a separate ngram-table.safetensors sidecar that MTPLX streams from SSD. The vision tower is preserved in the weights.

Bare Speed is every MoE expert and dense matrix at 4-bit with 64-weight groups, nothing promoted, with the same 16-bit set and the same n-gram sidecar. Both packs carry their sampling contract (temperature 1.0, top-p 0.95, top-k 20, the official Qwen 3.8 contract) in mtplx_runtime.json.

Exactness

Speculation in MTPLX is exact. Drafts from the MTP head are accepted with the probability-ratio rule and rejected drafts are resampled from the residual, so what you sample is what the model would have sampled without speculation, at any temperature. The draft sampler is a speed knob only. The quantization is the one approximation.

← Qwen 3.8 27B Qwen 3.6 27B →