Qwen 3.8 Next (Flash-Next) — MLX 4-bit with embedded MTP

An MLX-quantized version of Qwen3.8-Flash-Next for compatible Apple silicon runtimes, with native Qwen4 multimodal support and an embedded multi-token prediction (MTP) head. The text model, vision tower, learned n-gram tables and MTP weights are included in one indexed checkpoint. No separate draft-model download is required.

Artifact

The output projection uses 8-bit storage. Other per-module precision choices, scales and offsets are recorded in config.json; the 4-bit label does not mean every array uses four bits. Disk size is not a minimum-RAM specification: runtime buffers, context length, batching and offload policy affect memory requirements.

Multimodal support

The full vision tower, tokenizer, chat template, image processor and video processor configuration are included. This is a native Qwen4 vision-language checkpoint, not a text-only export or a Qwen3.5 architecture alias. Use a runtime that explicitly supports qwen4_exp and its native multimodal preprocessing. Audio support is not claimed.

The learned n-gram/PLE tables are retained. Compatible runtimes can keep these tables on SSD without changing their stored values. SSD offload, paged attention and prefix caching are runtime features, not automatic consequences of downloading the checkpoint.

Embedded MTP

The checkpoint includes one trained MTP prediction layer alongside the target model. An MTP-aware runtime can propose tokens with this layer and verify them against the target model before accepting output. Speculative depth is a runtime setting, not the number of trained heads or a change to the weights.

Embedding the head does not enable speculation by itself. Check the runtime's proposal and acceptance diagnostics. In the assessed d-inference implementation, eligible single-request text generation can use MTP; constrained required/named tool calls, media requests and multirow decoding use target-only paths under the existing compatibility gates.

Download and runtime compatibility

Download the complete repository with huggingface_hub:

from huggingface_hub import snapshot_download

model_path = snapshot_download(
    repo_id="EigenLabs/Qwen3.8-Flash-Next-MLX-4bit-mtp",
    local_dir="Qwen3.8-Flash-Next-MLX-4bit-mtp",
)

For reproducible deployments, supply a full immutable commit as revision. Load the resulting directory through a runtime with native Qwen4 support; generic MLX model loading does not establish architecture, multimodal or MTP compatibility.

Darkbloom integration is tracked in mlx-swift-lm #149 and d-inference #1030. These are draft integration PRs, not a claim that every shipping provider already supports the model. The native serving identity is DarkBloom/Qwen3.8-Flash-Next-Q4-mtp. Availability through a provider catalog depends on a separately approved runtime and immutable artifact revision.

Validation and limitations

The complete package was checked for index coverage, embedded MTP, vision and PLE inventories, and full per-file SHA-256 integrity. Local d-inference checks on these weights passed 59 standard compatibility cases and 24 string-tool and generated-tool-ID history cases in each MTP OFF/AUTO configuration, covering reasoning OFF/ON and streamed/nonstreamed Chat/Responses. The tested 81K context and uncached lifecycle cases also passed.

Known remaining failures include some adversarial literal-string copying and reasoning/tool cases, brief-video recognition, and related multimodal tool history cases. They are not established as exclusively model or engine defects. The trained tokenizer uses NFC Unicode normalization, so byte-exact preservation of decomposed Unicode input is not guaranteed.

These results do not establish BF16 equivalence, every cache or concurrency configuration, all physical hardware tiers, complete native state/MTP qualification, signed/trusted deployment, or hosted OpenRouter certification. Validate the intended workload and runtime before production use.

Attribution

Original model and training: Qwen. EigenLabs packaging and Darkbloom integration retain the supplied model license. This is not an official Qwen release.

See the original model card for the upstream model's intended use and limitations, and artifact-provenance.json and SHA256SUMS.json for package provenance and file hashes.