0.13.0 — Ornith-1.5, and prefill that tunes itself

0.13.0 brings the Ornith-1.5 family (9B dense + 35B-A3B MoE) and Nemotron-Labs-Diffusion 3B into the catalog, adopts MLX 0.32.1, auto-tunes recurrent prefill on verified profiles, and hardens the tool loop. The desktop app redesigns audio into two lanes, polishes the DMG install, and wires the first-run wizard onto the RAM-recommendation SSOT.

• New models. Ornith-1.5 — an open, MIT-licensed family — lands as first-class aliases: ornith-1.5-9b-bf16 (dense) and ornith-1.5-35b-a3b-bf16 (MoE A3B), both on the Qwen3.5-class GatedDeltaNet architecture with hermes tool calling (#2197). Nemotron-Labs-Diffusion 3B is served in AR mode as nemotron-labs-diffusion-3b-4bit (#2194). Qwen3-Next models can stream weights from disk (#2192).
• MLX 0.32.1. Adopted after Mac model benchmarks (#2199), with a guard that prevents quantized-KV corruption on the new runtime (#2234) and persisted KV isolated by model revision (#2202).
• Prefill tunes itself. Recurrent-model prompt prefill picks its chunk size from measured profiles instead of one fixed default, scoped to verified model profiles (#2210, #2211, #2213), and multimodal serving separates the MLLM prefill budget from the vision budget (#2215).
• A sturdier tool loop. Malformed tool calls are recovered through the tool loop instead of dropped (#2278), chat templates resolve from model profiles at load time (#2288), Aider stays pinned to the served local model, and Hermes gets cross-version toolsets.
• Desktop: audio in two lanes. Dictation and voice notes get separate readiness UX with the download cliff surfaced up front (#2188); the DMG install flow is polished (#2214); the first-run wizard picks models from the same RAM-recommendation SSOT as the CLI (#2232).
• Desktop: agents before the engine. “Connect your agents” is useful before an engine starts (#2308), agent health is surfaced with a re-run setup path (#2230), and the tray menu gains “Copy API endpoint” (#2246).