Red Lite dev42 — Qwen3-Coder-Next

Status: done on branch dev/coder-next (from main at a9193b9). Apple M4 Max 48 GiB only.

Why#

Qwen/Qwen3-Coder-Next (Apache-2.0, 2026-02-03) is a coding model with the Qwen3-Next-80B-A3B architecture: Qwen3NextForCausalLM, 48 layers, hidden 2048, 512 experts top-10, the same attention and DeltaNet shapes, and the same vocabulary. The only configuration difference is rope_theta (5e6 instead of 1e7), and it has no MTP tensors. SWE-bench Verified is 70.6 according to Qwen. Bartowski publishes it at exactly our file sizes. See docs/RESEARCH_2026_10.md.

Change#

None in the engine: the rope base is read from qwen3next.rope.freq_base. redlite-engine info prints base=5000000 for both files. redlite download gains the variants iq3_xxs (alias 48gb, the 0.4.0 file that was missing from the catalog), coder_iq2_xxs (alias coder) and coder_iq3_xxs (alias coder-48gb). Coder-Next is not picked automatically: it is a specialised model and must be passed by path.

Validation (M4 Max 48 GiB)#

file	bytes	SHA-256 (= Hugging Face LFS id)	dequantization vs ggml	`regress_m4.sh` vs pinned llama.cpp
`Qwen_Qwen3-Coder-Next-IQ2_XXS.gguf`	19,298,975,424	`992dcf62…483446`	bit-identical	50 / 50
`Qwen_Qwen3-Coder-Next-IQ3_XXS.gguf`	31,726,712,512	`0b0a3356…aaaf96`	bit-identical	37 passed, 0 failed, 13 skipped

The suite compares every native result with the pinned llama.cpp on the same file. That includes logits, greedy decode, 1200-token long context, full residency, server stream, restart from a stored state, and the sanitizer chat turn. The 13 IQ3_XXS skips are the IQ2_XXS-layout stage tools, as for the Instruct IQ3_XXS file. Full residency is 17,316 / 28,800 MiB, the same as the Instruct files.

Scope boundary#

Correctness only. Speed was not measured separately: the shapes and quant types are those of the Instruct files, but no benchmark was run on these files.
No quality evaluation of the model itself (SWE-bench or similar) was run here.
Not measured on the M4 Pro 24 GiB.