majentik/Swift-Qwen3.8-27B-MLX-8bit

Swift-Qwen3.8-27B-MLX-8bit

MLX 8bit (affine, group size 64) quantized variant of ukisai/Swift-Qwen3.8-27b (text tower quantized; vision tower and projector retained in BF16) for Apple silicon via mlx-lm.

Provenance

• —Source: ukisai/Swift-Qwen3.8-27b @ revision 048328f4059015b63f860a453bf94834af0db683 (Swift Open License v1.0 for the UkisAI contribution + Apache-2.0 for the Qwen base (upstream LICENSE, LICENSE-APACHE-2.0, NOTICE)).
• —Quantized with mlx_lm.convert (mlx-lm 0.31.3): affine, 8-bit, group size 64.

Smoke gate

Before upload this pack passed a deterministic coherence gate: greedy 48-token chat generation loaded through mlx_lm.load, judged for emptiness, repetition loops, multi-script gibberish, and special-token debris. Verdict: ok.

Usage

pip install mlx-lm
mlx_lm.generate --model majentik/Swift-Qwen3.8-27B-MLX-8bit --prompt "Hello"

Evaluation

Coherence smoke gate only at publish time (see Smoke gate above). Per-tier gate records live in majentik/garden-quant-bench; perplexity / benchmark rows are added here when measured.

License

This pack is a quantized (Object-form) redistribution of ukisai/Swift-Qwen3.8-27b and inherits its dual licensing unchanged:

• —the Qwen/Qwen3.8-27B base weights remain under Apache-2.0 (LICENSE-APACHE-2.0);
• —the UkisAI Swift contribution (finetune) is under the Swift Open License v1.0 (LICENSE), which is free for research and commercial use by organisations with up to US$1M gross annual revenue; larger organisations need a Swift Enterprise License from UkisAI.

LICENSE, LICENSE-APACHE-2.0 and NOTICE are shipped verbatim in this repository as required by Section 4 (Redistribution) of the Swift Open License. The only change from upstream is the MLX quantization described above.

Available tiers