The flagship Winter-Mix quantization of Qwen/Qwen3.5-122B-A10B for Apple Silicon: the best-measuring MLX quantization of this base model we are aware of, at any size — including against 94–95 GiB 6-bit-expert builds.

Native MLX, no strings attached. Standard MLX affine layout — bit-identical packing to mx.quantize, no custom kernels, no forked runtime. Loads in mlx-vlm, mlx-lm-based servers, LM Studio, and any downstream MLX app at exactly the speed of a stock-converted model. MoE router and vision tower are unquantized (BF16); MTP weights are stripped.

One scoring rule for every row — NLL over the second half of each context window, token-aligned across engines (llama.cpp's native counting rule) — paired per-token where both models run under MLX. Comparison rows are benchmarked reference points measured on the same harness, not published artifacts.

Best MLX numbers at both context lengths, 12–13 GiB smaller than the 6-bit builds it beats. In the extreme-depth band (49–98K positions) it is the strongest model we measured at any size. The imatrix-rounded source GGUF remains slightly ahead (+0.3–0.7% rule-matched) — matching it in MLX would require custom inference kernels, which this method excludes by design.

Out-of-domain (code + GSM8K, held out from all calibration): within ±0.1% of RTN references — no calibration binding. Multilingual (held-out 8-script eval): parity with English-calibrated builds.

Quantization can change how a model processes, not just how it scores — and perplexity is provably blind to some of it. This release was additionally selected on long-context behavioral testing: seeded state-tracking benchmarks scored blind, plus direct reading of the reasoning traces. WinterMix58 delivers perfect scores on a 30-step adversarial state-tracking task (every table row, every final state, every question, across all seeds), and its traces show disciplined self-verification — scheduled audit passes that resolve cleanly and twice caught a base-model reasoning bias before it reached the output.

Winter-Mix: sensitivity-informed mixed-precision allocation with hard pins on routing-critical tensors, calibrated error-compensated rounding (GPTQ-family, independent MLX-native implementation) executed layer-wise so a 122B quantizes on a single 128 GB Mac, and a diverse long-context calibration mixture engineered for full expert coverage — including multilingual content so language-specialist experts are calibrated rather than silently degraded. Developed and validated through 18 measured variants with paired controls. The full pipeline is not released.

Built by WinterCharm on a single M5 Max MacBook Pro. Taking requests for MLX quantizations of other models — open an issue in the Community tab.