• Two new models. GLM-5.3-Flash — 320B, and the first GLM that runs entirely in memory on a 128 GB Mac rather than streaming from disk. Qwen3.8-Flash-Next — 125B, the fastest large model in the app. Measured on an M1 Ultra: 24 and 31 tokens/sec respectively. • GLM-5.2 retired. At 217 GB it never fit in memory on a 128 GB Mac and only ever streamed; GLM-5.3-Flash replaces it and runs resident on the same hardware.

- Allow serving with license keys - Share your delight to family, friends, or potential customers - Added a few more models: - DeepSeek V4 Flash 0731, which claims to be the most cost-effective agent-aware LLMs from DeepSeek, that outperforms DeepSeek V4 Pro preview - Muse Glimmer 30B by Meta, which claims to be the most efficient agent-aware LLMs from Meta - Qwen3.8 27B, which claims to be the most efficient multi-modal LLMs that outperforms Qwen3.7 Plus models - MinMax H3, which claims to be a Visual LM that is capable of generate videos from a few snapshots and description - For agentic models, we added support for: - DeepSeek Harness 0.1.4 (dsh), the preview coding agent officially supported by Deep Seek that adopts composable programming plugins architecture - Grok Build, another Rust-friendly coding agent added - Engine updates: - Our DeepSeek V4 Flash model reference get DSpark integrated to gain up to 1.6x further speedup - Our Rust-based GGMF engine could perform as well as the llama.cpp on Apple Silicon now - Added a new Generate Video screen for: - The MinMax H3 model to be used to generate a 5~10s length video from natural language description, or two reference pictures (start/finish frames)

+ feat: adding support of fully-blown DeepSeek V4 Flash 0731 (mxfp4), this high quality model can run at half speed on 128GB machines compared to the quantized versions + feat: adding support of Laguna XS 2.1 model (it fits perfectly well in the memory of a 64GB machine) + feat: adding support of Kimi K3 model (quant 1bit, but it is the first model we serve that cannot run inside the memory of any existing single Mac) + feat: support long prefills to keep service alive, useful for the slow but high quality models to be on their jobs + feat: replaceable kernel hot patch with same ABI, useful for quick upgrade of the kernel efficiency without waiting for the next major release of the app + feat: enhancing ssd-streaming with double-buffered chunk pool and the per-layer decode chain + fix: loading problem to serve some models + fix: early stop of Laguna text generation

fix: code sign of the underlying serving component