Red Lite dev54 — a better expert mix at the same size

Status: on branch dev54/quant-mix. Measured on the Apple M4 Max 48 GiB, 2026-10-03/04. Roadmap Phase 4, "a Red Lite quantization".

The question#

Bartowski's IQ2_XXS file (the 24 GiB reference) keeps 11 of the 48 expert layers at IQ2_XS (2.31 bits per weight): layers 0–5 and 43–47. The other 37 are at IQ1_M (1.75 bits). Is there a better choice of those 11 layers at the same size?

Only the expert layers change. The native engine runs IQ2_XS and IQ1_M experts at this size; IQ2_XXS experts and mixed IQ1_M/IQ2_XS layers are not supported (see E1).

What the importance matrix says#

Bartowski's imatrix.gguf records each expert's input energy (sum of squared inputs per token). Weighted by how often each expert is used, the energy grows steadily with depth:

• down-projection input: 1.3 at layer 0, 9–13 in the middle, then 78, 109, 157, 246, 337, 490 and 565 from layer 41 to 47;
• gate input: 156 to 2001.

Bartowski's scheme spends half of its precise layers on the six with the lowest energy (0–5).

Method#

1. Source. Bartowski's Q8_0 (three splits, 84.8 GB) and his imatrix.
2. Quantizer. The pinned llama.cpp's llama-quantize --allow-requantize --imatrix --tensor-type-file. Every non-expert tensor gets exactly the type it has in Bartowski's IQ2_XXS (482 overrides). The pinned llama.cpp's own IQ2_XXS recipe differs (IQ2_XXS experts, Q4_K attention) and was not used. Each variant takes 37–43 minutes.
3. Variants:
   • R: Bartowski's scheme, reproduced;
   • E3: IQ2_XS on layers 37–47, the highest energy;
   • E3b: IQ2_XS on 0–2 and 40–47;
   • E1: IQ2_XS on every down projection, gate/up IQ1_M (+3 % size).
4. Measures:
   • perplexity on the frozen corpus (perplexity.sh, ctx 512, 111 chunks), with a paired per-chunk test against R: the per-chunk loss comes from the running estimates in llama-perplexity's log;
   • agreement with Qwen's own API (api_compare.py, 235 prompts, every expert resident).

Results (M4 Max 48 GiB)#

• E3 is better than Bartowski's scheme at the same size, on both measures. Perplexity is −0.94 %, and the paired test is clear (t = −3.9; it wins on 72 of 111 chunks). It also gets more identical answers and more matching words against the API.
• E3b (keeping three early layers precise) is not distinguishable from R.
• E1 is worse and 3 % larger.
• Correctness. E3 passes regress_m4.sh 52/0/0 and quick_parity.sh (CPU vs Metal, GPU-routed, logits and greedy vs the pinned llama.cpp on its own references).
• Speed. Decode is unchanged: same types, same footprint (17,316 MiB at full residency; 87.5 tok/s on the sky-is-blue prompt).
• R vs Bartowski's own file. R measures 16.370 against 16.4675 for Bartowski's file (dev31) with the same corpus and perplexity build. The quantizer version differs; the variants are compared with R, built the same way.

Reproducing#

python3 scripts/dev/quant_mix.py --like models/Qwen_Qwen3-Next-80B-A3B-Instruct-IQ2_XXS.gguf \
    --q8 models/q8/Qwen_Qwen3-Next-80B-A3B-Instruct-Q8_0-00001-of-00003.gguf \
    --imatrix models/q8/Qwen_Qwen3-Next-80B-A3B-Instruct-imatrix.gguf --iq2xs-layers 37-47 --out E3.gguf

It needs the Q8_0 (84.8 GB download), the imatrix, the pinned llama.cpp (make bootstrap) and about 40 minutes.

Scope boundary#

• Only one size (≈ 19.3 GB) and one model (Qwen3-Next-80B-A3B-Instruct) were studied. Four variants were compared; this is not a search over all layer sets.
• Quality is measured by perplexity on one corpus and agreement with Qwen's API, not by task benchmarks.
• E3 is published (dev54b) at alfodaniello/Qwen3-Next-80B-A3B-Instruct-RedLite-GGUF:
  • SHA-256 b62067d6f28c52c07f8fca55196e8a63916264e40c6641f6a8d2e710768d291c;
  • redlite download 24gb (alias e3) fetches it, and redlite chat prefers it over Bartowski's file;
  • on the M4 Pro 24 GiB the download took 6 minutes and the hash matched; the sky-is-blue answer is right, at 26.7 tok/s with the 4 GiB cache.