0.13.2 — Flash-Next decodes faster, the app gets 43% smaller

Released 2026-08-30 · full changelog · GitHub releases

Long sessions get quicker on both ends: Qwen3.8-Flash-Next can opt into its own multi-token prediction head, long prompts reach the first token roughly a third sooner, and a repeated prompt reuses its cache instead of paying for it twice. The Mac app ships as a download less than half the size it was, Kokoro is ready before you go offline, and attachments, credentials and model switching got a safety pass.

• Native MTP for Qwen3.8-Flash-Next. Opt in and the checkpoint's own one-layer prediction head proposes tokens while the full model verifies every one of them, so the output is the model's — not an approximation. Measured on an M3 Ultra: 25.2 → 34.9 tok/s at a short prompt and 21.2 → 28.8 tok/s at 32K (+36–42%), with 76.41% of proposals accepted. It costs up to 6.6 GB more active memory, and seeded requests or stateful grammar, tool and reasoning processors deliberately stay on ordinary decoding rather than quietly weakening their contract. 192 GB remains the recommended tier for this experimental checkpoint (#2572, #2655).
• Long prompts reach the first token ~a third sooner. QSA index-cache work is now batched across eligible prefills: 2K 3.35 → 2.26 s, 8K 13.69 → 9.24 s, 32K 62.85 → 44.66 s (−29–32%), with decode speed and cache precision unchanged (#2574, #2596).
• Repeated prompts get their cache back. Prefix snapshots now follow the exact rendered prompt and survive batching and persistence, so a warm 5,288-token request reused 5,273 tokens and finished in 0.539 s — and a completed 32K request can no longer poison the next one in the same process (#2588, #2644).
• Kokoro is ready before you go offline. rapid-mlx pull now fetches the voice assets and prepares the English G2P requirement as part of the pull — validated by generating speech with networking disabled after a clean cache. A missing runtime requirement returns an actionable readiness error instead of installing software mid-request (#2648, #2664).
• The Mac app download is 43% smaller. LZMA packaging and dependency-proven pruning took the signed, notarized DMG from 187.5 MB to 106.3 MB. The honest tradeoff: the one-time Finder copy is slower on the measured host, 5.94 s → 17.80 s. Releases now publish the exact approved and notarized bytes rather than rebuilding after the tag (#2668, #2775).
• Attachments, credentials and model switching got a safety pass. A message takes up to four images and 6 MiB of encoded image data and says which limit rejected the rest; deleting a generated image uses an app-owned sheet where Return cannot delete; switching away from a busy model asks first and Cancel keeps the in-flight response; and Tools settings resolve a search key's real Keychain state without ever exposing it (#2541, #2578, #2543, #2514).
• Tool calls and explicit flags mean what they say. Qwen3.8-27B required and named tool calls use the checkpoint-native format and return OpenAI-compatible JSON arguments, an invalid required argument now fails as a client error instead of looking executable, explicit --mllm wins over automatic fallback, and an explicit variant pull persists its subfolder so a later serve resolves the checkpoint you actually pulled (#2660, #2643, #2558).