Qwen3.5-122B-A10B-wMix38

Qwen3.5-122B-A10B — WinterMix38 (MLX)

59 GiB · 4.12 bpw measured · 3-bit gate/up + 4-bit down expert core / 8-bit backbone · vision intact · native MLX format.

The most memory-lean Winter-Mix quantization of Qwen/Qwen3.5-122B-A10B — and, at long context, the best one. At 16K context this 59 GiB build posts a better perplexity than our 82 GiB flagship and the 85.6 GiB source GGUF it descends from, and on held-out agentic session data it outscores every released WinterMix, thanks to the new annealing process on its reasoning traces (new at the wMix38 tier, not applied to wMix48, wMix58, or wMix68 yet). On a 128 GB machine it leaves ~45 GB free for parallel long-context agent sessions (this architecture's GDN-dominant design keeps a 100K session's cache at only ~5–10 GB; raise iogpu.wired_limit_mb for swarm use).

Native MLX, no strings attached. Standard MLX affine layout — no custom kernels, no forked runtime, nothing the runtime has to know about; loads anywhere MLX models load, at stock speed. Router and vision tower BF16; MTP stripped.

Method (brief)

Winter-Mix Quantization: sensitivity-informed mixed-precision allocation with hard pins on routing-critical tensors, GPTQ-family rounding driven by a diverse long-context calibration mixture engineered for full expert coverage — and, new at this tier, a final stage that anneals the reasoning chains on the model's own running token-stream statistics. That last stage is what turns the 3-bit tier from a compromise into the family's long-context specialist. It was developed across a seven-variant controlled program with pre-registered predictions, behavioral probes, and full trace reads at every step. The full pipeline is not released, as the method is still in active development. This is the most coherent 3-bit quantization of Qwen3.5-122B-A10B available in native MLX format.

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

Measured quality

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). Comparison rows are benchmarked reference points measured on the same harness, not published artifacts.

• At 16K context it beats every MLX build and the source GGUF — the flagship by −1.9%, the 85.6 GiB Q5_K_XL by −1.2%, oQ4 by −5.4% — from near the smallest footprint in the table. The quantization tax lives at short context (+2.5% vs the flagship) and inverts exactly where long-session work lives.

• Against the best imatrix 3-bit (Unsloth's UD-Q3_K_XL): a statistical tie up to 16K — and past it WinterMix38 dominates.

  Extended head-to-head, same token stream, both engines, same rule, here's the data:

  context	UD-Q3_K_XL	WinterMix38	margin
  2K	4.3460	4.3556	imatrix +0.22%
  16K	4.3264	4.3329	imatrix +0.15%
  24K	4.7728	4.6666	WinterMix38 −2.2%
  48K	5.6064	5.4232	WinterMix38 −3.3%
  96K	5.5910	5.2733	WinterMix38 −5.7%

  The crossover sits at ~16–20K, and the margin grows monotonically with depth — at 96K every individual window favors WinterMix38. On the 5-bit tier the imatrix format held a +0.3–0.7% edge over native MLX everywhere; at 3-bit this method holds it to ~0.2% at short context and beats it by 2–6% everywhere long-session work actually lives — without imatrix weighting, without custom kernels, in the format your Mac already runs.

• On held-out agentic session data it outscores every released WinterMix — 2.7385 vs the flagship's 2.7627, WinterMix48's 2.7590, and WinterMix68's 2.7594, same harness, same rule.

• Long-context verified to 98K with no depth collapse; in the 98K band it holds the best number of the seven-variant controlled program that produced it, by a wide margin. Multilingual (8-script held-out) eval: tied-best of its tier.

Beyond perplexity

Selected and verified with the same three-instrument protocol as the rest of the family — NLL, seeded behavioral probes, and direct reading of every reasoning trace. On a 30-step adversarial state-tracking benchmark (5 seeded runs, 50K-token context), WinterMix38 delivered the exact correct final state, total, and spoilage accounting in 5 of 5 runs — zero fabricated figures, zero truncations, and not a single cell deviation across 750 tracked state values. A texture note from the trace reads: this build executes its working steps in clean single passes and concentrates its self-verification at the answer stage — where in three of five runs its audits caught and reversed a drafted comparison error before delivery. Measured generation speed: ~35 tok/s at 50K context on an M5 Max.

Next Steps

A 2-bit quantization of Qwen3.5-122B-A10B on MLX is in the works — the WinterMix28 release of this model. The 3-bit work translates broadly to 2-bit, and the pressure of the 2-bit tier is refining the method further; it ships when it passes the same gates as everything else in this family — pre-registered predictions, behavioral probes, and full trace reads. After that, I'll be applying the "reasoning annealing" upgrade to the previously released wMix58 and wMix48 quantizations as well — it's a genuinely new technique I had to invent to make the wMix38 quantization possible, and every WinterMix deserves the most effective form of it.

Which one?

• WinterMix58 (82 GiB) — the flagship: best short-context quality, the default for interactive chat and mixed workloads.
• WinterMix48 (68 GiB) — the balanced middle option.
• WinterMix38 (59 GiB, this model) — the long-context specialist: best 16K and extreme-depth quality of the whole family, plus the deepest memory headroom. If your sessions run 16K+ tokens — agents, long documents, deep multi-turn work — this is the strongest 122B build we make, at any size.
• WinterMix68 (95 GiB) — maximum per-weight fidelity, calibration-free reference.

Credits

• Base model: Qwen/Qwen3.5-122B-A10B (Apache 2.0, inherited).
• Allocation informed by analysis of the tensor-type metadata in Unsloth's published Dynamic GGUFs (metadata analysis only; no weights or code reused).
• GPTQ: Frantar et al., arXiv:2210.17323.
• Built on MLX / mlx-vlm.