Qwen3.8-Flash-Next MLX Activation-Weighted 3-bit

This is a reproducible MLX conversion of Qwen/Qwen3.8-Flash-Next at immutable revision f5d08274bafd880402bd16f5e3e6c514136ec06c.

The 48 base routed-expert banks use fresh MLX affine Q3/group-64 codes generated directly from the original BF16 checkpoint. Their scales and biases are refit against frozen image-and-text expert-input second moments. Remaining eligible core, MTP, and vision matrices stay Q4; the 160-wide n-gram table stays Q4/group-32.

• Artifact payload: 83.49 GiB
• Quantized Q2 modules: 0
• Quantized Q3/group-64 modules: 144
• Quantized Q4/group-32 modules: 128
• Quantized Q4/group-64 modules: 783
• Source: 180B parameters including 125B main, 51B n-gram embedding, and 4B MTP
• Native context: 262,144 tokens

Run locally with mere.run

This profile requires the managed model update after mere.run v0.45.0. With a build that contains that update, review the license and pull the checkpoint:

mere.run model pull vision-chat-q38-flash-next-3bit \
    --accept-license-terms

The repository is public and ungated. The acceptance option records your acknowledgment of the bundled Qwen Community License 1.0.

To generate text with the bundled multi-token prediction (MTP) head, run:

mere.run text chat \
    --model vision-chat-q38-flash-next-3bit \
    --context-size 32768 \
    --max-tokens 256 \
    --temperature 0 \
    --no-thinking \
    --stream \
    --stats \
    --prompt "Explain sparse attention in three short sentences."

For an external SSD, add --cache-dir /Volumes/Models/huggingface-cache to the pull command and replace the sample path with your mounted volume. Keep the volume mounted when you use the model.

Native qualification

The fresh Q3 checkpoint passed a no-regression comparison against the pinned published Q4 checkpoint in mere.run on a 128 GiB Apple Silicon Mac. MTP was disabled during checkpoint selection.

• The candidate passed 58 of 61 cases by exact expected output.
• The three absolute misses matched the published Q4 output exactly.
• The sealed holdout produced identical candidate and Q4 outputs in all 16 cases, including exact output on all eight image and OCR cases.
• The holdout peak memory footprint was 62,065,713,472 bytes for this profile and 77,180,432,496 bytes for Q4, a 19.58% reduction. Neither run increased swap usage.

The bounded suites don't establish general model quality or BF16 parity. For the test identities, hashes, memory measurements, and limits, see MERERUN_QUALIFICATION.json.

Runtime status

The tensor inventory, source hashes, MLX packing, fused-expert split, convolution layout, and zero-centered RMSNorm conversion are validated by the bundled MERERUN_CONVERSION.json. Use a Qwen4Exp-aware runtime such as mere.run or mlx-vlm.

License

This redistribution retains the upstream Qwen Community License 1.0 in LICENSE. Review it before use. In particular, it contains attribution/display requirements for very large commercial products and separate-license conditions for certain commercial Model-as-a-Service and AI Work Assistant uses. The model is not gated; downloading or using it does not remove those terms.

The upstream model card is preserved as README.upstream.md.