v0.1.5 — the judgment release
Recurrent and hybrid models no longer crash. Qwen3.5's delta-net layers (like Mamba and Jamba states) cannot roll back past positions, which speculative decoding's rejection step requires — so reame run died mid-generation with a cryptic "failed to truncate sequence" on the whole Qwen3.5 family. Reame now detects these architectures and downgrades to classic decoding automatically, logging why.
Qwen3.5-9B joins the catalog as the judgment pick. We measured four models on a real task — an SEO audit of a live webpage: a 1.5B and OLMoE invented findings (flagged missing alt text that exists), a 30B-A3B miscounted the meta description length, and the 9B got everything right, including an empty <h2> the page really has, in 73s on a laptop (16.6 tok/s plain decode, 5.3 GB). reame run qwen3.5-9b and you have it. The README gains a verified one-liner that audits any live page from its raw HTML.
Extraction stays the small models' turf — that's the measured boundary: when the answer lives in the context, run a 1.5B at 81 tok/s; when the task needs judgment, batch the 9B.
Prebuilt binaries: Linux x64, Linux ARM64, macOS ARM64 (runtime dependency: libzstd).