llama.cpp v0.6.0: faster Metal mat-mul, GLM-5.3-Flash, /v1/systemone

llama.cpp v0.6.0 was published on GitHub on 2026-10-05. It matches nightly build b11429 and is not a prerelease. Metal gets new few-row MMA mat-mul kernels for speculative and batched decoding. The release notes say these give "up to ~3x faster mat-mul on Apple GPUs" (#29869). The linked pull request says the kernels run on Apple GPUs without the tensor API (M1 to M4); M5 uses the tensor API by default. The ~3x figure is for mat-mul only, not end-to-end tokens/s, and we have not benchmarked it. Metal also adds a tensor-API flash attention kernel for F16 KV (#29570), which its pull request describes as a kernel for M5. Vulkan gains sparse flash attention for quantized K/V, and ggml moves to v0.26.0. Qwen4Exp now has MTP speculative decoding. The notes state a ~1.5x decode speedup on DGX Spark (#29761). New models: GLM-5.3-Flash (GLM5-Next), which the notes describe as a 320B KDA/DSA hybrid text+vision model with mHC and MoE (#27773). Also added: the Clef decision model (text and vision), Ling 3.0 VL, the Nimble decision model, and LFM2.5-Encoder-230M/350M. API: a new `llama_batch_ext` batch API with `llama_process()` is added. Session formats move to `LLAMA_SESSION_VERSION` 11 and `LLAMA_STATE_SEQ_VERSION` 4. Server: a new `/v1/systemone` endpoint serves decision models. `/v1/embeddings` accepts typed vision, audio and video content. `GET /v1/models` now reports each model's input and output modalities. The web UI adds a model download pipeline, a Hugging Face Hub data layer and model memory-fit estimation.