MLX Serve Files

• DeepSeek-V4.1-Flash across two Macs. --dist tensor splits the model over two Thunderbolt 5 Macs that work on every token together: on two M3 Ultras it writes 57% faster with DSpark (88 tokens a second instead of 56) and 31% faster without. Thanks @ivanfioravanti.
• DeepSeek-V4.1 GGUF files at MLX speed. With --mlx-gguf, DeepSeek-V4.1 GGUF sets (ISTA's GSQ-RCO packs, the Coder pack and ivanfioravanti's DS4 packs) run on the MLX engine with DSpark drafting and image input (a llama.cpp image projector beside the files works too): about 3,300 tokens a second reading prompts on an M5 Ultra, 2.5x what it managed before and close to our own MLX pack. More GGUF families take this path too (Qwen3, Flint, Muse Glimmer), with faster IQ kernels, and GGUFs on the built-in DeepSeek engine open at the context they were trained for instead of 32k (thanks @garnetlyx).
• Faster and steadier. DeepSeek-V4.1's Lossless pack writes 7% faster on M3 and older Macs (thanks @ivanfioravanti), its iQ-MLX and expert-pruned packs write 25% to 43% faster on M4 and older Macs, 3-bit mixture-of-experts models such as Qwen3-Coder-30B-A3B write a little faster (thanks @toxeh), and a model stays in memory between requests instead of being squeezed out by macOS, which could freeze a Mac running a model close to its memory size.
• Long chats stay cached when you go back. On models that keep their own state (DeepSeek-V4.1 and others), a conversation keeps a few earlier checkpoints, so editing or forking an older message answers in a fraction of a second instead of re-reading the whole chat (8.7 s to 0.13 s at 60k tokens). Thanks @STRML. Forking a long Qwen3.8-Flash-Next chat that moved to the SSD picks up from a checkpoint too: 5 seconds instead of 46 (thanks @yoyo930021).
• Newer MLX. mlx-serve moves to MLX 0.32.5. Outside plugins built for 26.11.1 now show Needs rebuild in Models ▸ Plugins, and one Rebuild all compiles them again with this release's kit; a model a refused plugin served names it in its load error. Outside plugins can now use the Neural Engine and stream their weights from the SSD with their own memory budget, and a server can steer a model per request by removing a direction from its activations (thanks @STRML, @sjordan0228).
• Answers and switches come out right. Switching away from a DeepSeek-V4.1 Sushi pack unloads it instead of refusing the new model, Qwen3.8-Flash-Next picks up its cached reply on the next agent turn instead of re-reading it, agents talking to DeepSeek's API no longer fail on tool-call history, GLM-5.3-Flash no longer carries one turn's reasoning into the next, a reply that keeps restating one paragraph with a reworded ending is cut short (thanks @lborloz), Qwen3.8-Flash-Next no longer crashes when an image request shares the GPU with other chats (thanks @slava-kudzinau, @yoyo930021), DeepSeek-V4.1 runs on Macs built with the macOS 26.5 tools (thanks @pcoppinger), headless Qwen3 embedding models load (thanks @trigger2k20), and a memory refusal with nothing left to unload says a restart is needed. A reasoning_effort word (low, medium) no longer cuts a model's thinking short: it reaches the chat template as in other engines, and only a reasoning budget you set caps it.
• App fixes. The model browser opens on a model you already have (thanks @trigger2k20), a GGUF picked in the app finds its own settings (thanks @elkaix), Models ▸ Plugins downloads the pack you clicked instead of the first one on the list (thanks @hyspace), and turning an MCP server on or off in the marketplace keeps the command, environment and folder you set by hand in mcp.json (thanks @jondestoppeleire).

Thank you

@elkaix, @fedeandri, @garnetlyx, @hyspace, @ivanfioravanti, @jondestoppeleire, @lborloz, @pcoppinger, @sjordan0228, @slava-kudzinau, @STRML, @toxeh, @trigger2k20, @yoyo930021