SuperQwen3.8-Flash-Next-abliterated-MLX-4bit

This is the measured MLX 4-bit deployment variant of Jiunsong/SuperQwen3.8-Flash-Next-abliterated@e7b3664cd9a5096a674db80f6ba3d379401cbc71. The derivative originates from Qwen/Qwen3.8-Flash-Next. It does not introduce a separate behavioral fine-tune; it converts the sealed SuperQwen payload using the immutable converter revision recorded below.

Variant

Format: MLX 4-bit
Quantization: {"bits": 4, "effective_bits_per_weight": 4.675, "exceptions": "PLE group-32 and small routing gates 8-bit as emitted by the pinned converter", "group_size": 64, "mode": "affine"}
Converter: Blaizzy/mlx-vlm@6102cb4ad1a5b3cc38d8dc7e6cbe2aca395596cb
Payload files: 31; payload bytes: 103691776762

Verified runtime

One real served inference call and the capability, tool-call, strict structured-output, reasoning, and formatting gates passed on Apple M5 Max, 128 GB unified memory, macOS using mlx-vlm.

C1 decode: 34.995 tok/s
C1 prefill: 169.811 tok/s
Directly tested context: 4143 tokens
Recipe: python -m mlx_vlm.generate --model . --prompt "Describe the runtime check."

These numbers apply only to the sealed hardware, runtime, command, and artifact in release_evidence/runtime.json.

Integrity

The conversion and served runtime reports both bind to payload identity 3697b0ed2ff4003581d3d4e90dad2c0e47506c749f22440b77b8786e1b6202fe. SHA256SUMS.json seals every published file. Private-revision, anonymous public, and immutable-tag verification is recorded separately by the publisher.

Limitations

Direct context validation is bounded to 4,143 prompt tokens; longer contexts are not claimed.
The first cold request measured about 67 seconds TTFT while the 97 GB payload paged in and compiled; warm C1 TTFT was 1.72-1.79 seconds.
Native MTP and external speculative draft decoding were not served in this MLX release and are not advertised.
The pinned MLX-VLM server logged a complete Uvicorn shutdown, then the process returned 139 during teardown; all served gates completed before shutdown and the artifact remained unchanged.

This structural derivative changes refusal behavior relative to the upstream model. No compatibility, context, speed, quality, or universal behavior claim is made beyond the included measurements. Review the upstream Qwen license and terms before use.

Attribution

Upstream: Qwen/Qwen3.8-Flash-Next
Behavioral derivative: Jiunsong/SuperQwen3.8-Flash-Next-abliterated@e7b3664cd9a5096a674db80f6ba3d379401cbc71

SuperQwen3.8-Flash-Next-abliterated-MLX-4bit

Variant

Format: MLX 4-bit

Quantization:

{"bits": 4, "effective_bits_per_weight": 4.675, "exceptions": "PLE group-32 and small routing gates 8-bit as emitted by the pinned converter", "group_size": 64, "mode": "affine"}

Converter: Blaizzy/mlx-vlm@6102cb4ad1a5b3cc38d8dc7e6cbe2aca395596cb

Payload files: 31; payload bytes: 103691776762

Verified runtime

One real served inference call and the capability, tool-call, strict structured-output, reasoning, and formatting gates passed on Apple M5 Max, 128 GB unified memory, macOS using mlx-vlm.

C1 decode: 34.995 tok/s

C1 prefill: 169.811 tok/s

Directly tested context: 4143 tokens

Recipe: python -m mlx_vlm.generate --model . --prompt "Describe the runtime check."

These numbers apply only to the sealed hardware, runtime, command, and artifact in release_evidence/runtime.json.

Limitations

Direct context validation is bounded to 4,143 prompt tokens; longer contexts are not claimed.

The first cold request measured about 67 seconds TTFT while the 97 GB payload paged in and compiled; warm C1 TTFT was 1.72-1.79 seconds.

Native MTP and external speculative draft decoding were not served in this MLX release and are not advertised.

The pinned MLX-VLM server logged a complete Uvicorn shutdown, then the process returned 139 during teardown; all served gates completed before shutdown and the artifact remained unchanged.