0.14.2 — Measured decode gains across four model families, and an approval-first Agent mode

Five qualified paths got faster on named hardware, and every one of them falls back to the stock route outside the shape it was tested on. Desktop gains an opt-in Agent mode that pauses before a consequential action and waits for you to approve the exact pending call. K2 Horizon 7B arrives on a Rapid-owned native runtime, and a set of prompt-cache and tool-loop failures found in real Desktop use are closed.

• Five measured speed-ups on the product path. DeepSeek V4.1 Flash on mixed DSpark K4 9.58 → 19.39 tok/s (2.02×); Qwen3.6 35B on the native MTP fixed suite 83.32 → 130.93 tok/s (+57.1%); GLM-5.3 Flash across six real tasks 26.58 → 35.54 tok/s (+33.7%); Qwen3.6 35B compiled no-MTP HTTP decode 101.7 → 121.6 tok/s (+19.6%); and its fused GDN HTTP path 62.3 → 79.1 tok/s (+27.0%). GLM-5.3 preserved all 12 reasoning traces and final answers byte-for-byte in its paired qualification. The Qwen3.6 kernels and the compiled replay fail closed to the stock route when the request shape, cache mode, concurrency or runtime is outside the tested set.
• Agent mode in Desktop, as an experiment. An opt-in mode backed by the same bounded runtime as Server: it can drive a local model through configured MCP tools, pause before a consequential action, show a bounded and credential-redacted summary, and continue only after you approve the exact pending call. The first version is deliberately conservative — runs are authenticated and process-local, completed state expires, tool sets are capped, unknown tools are treated as side effects, and an MCP reload cannot redirect an already approved call to a replacement tool. A denial records that the action did not execute rather than asking a compact model to reinterpret the refusal (#3480, #3439).
• K2 Horizon 7B on a native runtime, as an experiment. Server and Desktop run it through a Rapid-owned implementation including its reasoning and tool-call formats. On an Apple M4 Pro with 48 GB of unified memory the shipped paths measured 49–51 tok/s, 0.693 s median TTFT and 5.52 GB peak MLX memory — roughly Qwen3.5 9B decode speed at slightly lower peak memory, and slower and heavier than Qwen3.5 4B. The Rapid adapter and the checkpoint-bundled implementation produced the same output-token hash with median throughput 0.0089% apart, so no speculative performance patch was added. It stays experimental and is not a default recommendation (#3486, #3483).
• Long Desktop conversations stop paying to re-read themselves. A web-search or tool turn no longer moves temporary guidance to the front of the prompt, so a long conversation keeps its prefix cache instead of spending roughly 20 seconds re-reading itself on both the tool turn and its follow-up. Hybrid-model conversations resume from recurrent-state checkpoints, and exact prompt preparation is reused rather than repeated (#3454, #3446, #3435).
• Tool calls fail legibly instead of corrupting the edit. Malformed tool arguments now preserve the answer text already generated and give the model a precise correction target instead of surfacing raw protocol debris. An XML parameter value keeps the indentation of its first line, so a coding agent's replacement block still compiles — the parser was stripping all surrounding whitespace where exactly one newline per side is markup, which matches what vLLM and SGLang do for the same wire format (#3419, #3369). Document search, outlines, numbered sections and malformed retrieval cursors fail honestly rather than returning plausible but incorrect results.
• Code-editing turns reuse what is already on screen. Copy-draft now reaches sampled and cache-reusing requests: measured Qwen3.5 and Qwen3.8 turns accepted 71–99% of copied tokens and decoded 1.6–2× faster when the answer reused text already present in the conversation. Requests with nothing useful to copy are unchanged (#3417, #3398).
• Two more text aliases. k2-horizon-7b-4bit and deepseek-v41-flash-reap-2bit bring the catalog to 192 text aliases, 246 total. The DeepSeek REAP 2-bit checkpoint measured in 0.14.1 and held back now ships as an explicitly experimental alias behind a 224 GB minimum-memory gate — reachable by name on a machine that can hold it, still not a recommendation for ordinary Macs and still not a managed download.