Problem
Some newer model families are present in the cache but do not rank well because metadata or benchmark evidence is missing. Examples seen locally: Qwen/Qwen3.6-35B-A3B is not detected as MoE, and Qwen/Qwen3.5-9B / newer coder models have weak or missing benchmark evidence.
This also affects task profiles: --profile math is currently very sparse, and --profile coding still falls back to older Qwen2.5-Coder models quickly.
Done when
Qwen/Qwen3.6-35B-A3B is parsed with correct MoE/active-parameter metadata or an equivalent curated override.- Newer Qwen3.5/Qwen3.6 coder/math-relevant models have reasonable direct, inherited, or curated benchmark evidence.
--profile coding and --profile math return current, defensible recommendations on common simulated GPUs.- Tests cover the new metadata and ranking behavior.
Problem
Some newer model families are present in the cache but do not rank well because metadata or benchmark evidence is missing. Examples seen locally:
Qwen/Qwen3.6-35B-A3Bis not detected as MoE, andQwen/Qwen3.5-9B/ newer coder models have weak or missing benchmark evidence.This also affects task profiles:
--profile mathis currently very sparse, and--profile codingstill falls back to older Qwen2.5-Coder models quickly.Done when
Qwen/Qwen3.6-35B-A3Bis parsed with correct MoE/active-parameter metadata or an equivalent curated override.--profile codingand--profile mathreturn current, defensible recommendations on common simulated GPUs.