Qwen3.8-Flash-Next-DERISKED-MTPLX-Optimized-Speed

MTPLX.COM: 2 to 3x speedup. The fastest way to run models on a Mac.

Qwen 3.8 Flash-Next DERISKED Optimized Speed

Dynamic 4-bit quant with 8-bit attention. Higher quality and slightly slower. Built from the Blackfrost DERISKED checkpoint, for operator-controlled security research.

Blackfrost's DERISKED release of Qwen's 125B-A6B Flash-Next preview, the Qwen4-generation architecture with GDN hybrid MoE, Qwen Sparse Attention, and the 51B-parameter n-gram memory, running natively on MTPLX with its native multi-token-prediction head drafting through MTPLX's speculative lane. This is the same recipe as the first-party Optimized Speed pack: the Qwen Sparse Attention projections are kept at 8-bit, so the attention pathway that steers long contexts keeps its precision. For the unmodified Qwen checkpoint, pick that first-party pack instead.

The 32 GB n-gram embedding table streams from SSD by default, so the model fits a 96 GB+ Apple Silicon Mac with headroom. The weights stay resident, the table does not have to.

What DERISKED means

The base checkpoint is Blackfrost's, not Qwen's. Per Blackfrost's model card, its refusal surface was modified at the weight level for operator-controlled security research and red-team evaluation; no SFT, DPO, LoRA, merging, expert pruning or quantization was applied upstream. In Blackfrost's words: "This is not a safety-stock checkpoint and must not be deployed, marketed, or evaluated as one." This pack inherits that status. It changes the numeric format, not the behavior.

The chat template is copied verbatim from the Blackfrost checkpoint and carries the Blackfrost Qwentium operational prompt: a behavioral instruction, not an authorization system or a security boundary. Safety controls belong outside the model. Blackfrost recommends authenticated inference endpoints, independent request and tool-execution logging, least-privilege credentials, sandboxed code and file access, and explicit human approval for irreversible actions.

Speeds

Measured on an M5 Max with 128 GB, single stream, real server (mtplx serve, Turbo profile, MTPLX 2.10.2), official Qwen 3.8 sampling (temperature 1.0, top-p 0.95, top-k 20; sampled, not greedy).

That is a 1.4x speculative multiplier through the product serve path, on sampled output that follows the model's own distribution.

How it is built

• MoE experts and dense matrices at 4-bit with 32-weight groups; the Qwen Sparse Attention projections, the GDN output projection, the MoE router, the shared expert, the QSA indexer, the embeddings and the output head promoted to 8-bit (the quality edge over a flat 4-bit build).
• The GDN convolution and recurrent-state parameters, the hyper-connection mixers, the PLE projections and every norm stay 16-bit. The MTP head ships as mtp.safetensors with its experts at 4-bit.
• The n-gram embedding table ships as a separate ngram-table.safetensors sidecar (4-bit) that MTPLX streams from SSD (resident is opt-in on very large machines). The vision tower is preserved in the weights.

The serving contract ships inside mtplx_runtime.json. MTPLX reads it on load. Drafts are accepted with the probability-ratio rule plus residual resampling, so the output follows the model's own distribution at any temperature.

Use it

Mac app: download at mtplx.com, then find this pack under Forge, Discover (search the repo name).

Command line:

pip install mtplx
mtplx serve --model p0ly31/Qwen3.8-Flash-Next-DERISKED-MTPLX-Optimized-Speed

Siblings, built from the unmodified Qwen checkpoint: Optimized Speed (same recipe) and Bare Speed (flat 4-bit, the quickest build).

Base model: Blackfrost-AI/Qwen3.8-Flash-Next-DERISKED-BF16 (Qwen Community License; the Blackfrost model card is preserved in this repo as README-upstream-blackfrost.md). Upstream: Qwen/Qwen3.8-Flash-Next.