llama.cpp ships v0.5.0.
llama.cpp v0.5.0 GA — Metal MoE fusion, multi-bind, pin-able release
Release highlights: CUDA conv2d implicit GEMM; Metal MoE and SSM_CONV fusion (#28948); server bind to multiple addresses (#28690); input_image on function-call outputs; HRM-Text / MiMo-V2.6 / HunyuanOCR conversion support; router lifecycle fixes; cpp-httplib 0.57.1. Nightly note on the GA page: build b11146.
Why it matters: operators finally get a pin-able GA (not only nightlies) with Apple Silicon Metal MoE fusion and multi-bind for llama-server — prefer v0.5.0 unless you need tip-only features (tip now b11158).
Source: v0.5.0
None this window.
| Model / package | Org | Notes |
|---|---|---|
| llama.cpp v0.5.0 ~4:50 PM ET Sep 23 |
ggml | GA Metal MoE / SSM_CONV · multi-bind · ggml 0.25.0 |
| Ollama v0.34.4 assets ~12:18 AM ET Sep 24 |
Ollama | GA MLX Qwen 3.8 PP · Gemma 4 dyn res |
| llama.cpp tip b11158 ~7:40 AM ET Sep 24 |
ggml | Tip Vulkan Adreno coopmat · Ling 3.0 VL in b11156 |
| Gemini 3.8 Flash TTS Sep 23 |
API gemini-3.8-flash-tts / flash-lite-tts |
|
| ExecuTorch 1.5.1 ~12:58 PM ET Sep 23 |
PyTorch | GA patch; Core ML/MLX C++ SDK = 1.5.0 |
| Qwen3.8-Flash-Next-4bit-mtp ~4:48 PM ET Sep 23 |
mlx-community | Hub + LensVLM-9B-OptiQ-4bit · not mlx GA |
| Item | Beat | Status |
|---|---|---|
| vLLM Sep 23 GHSA quartet ~6:16–6:25 AM ET Sep 23 |
EngineCore kill · GLMGA video · Rust Prometheus · Qwen-VL max_frames — patched ≥ 0.30.0 | Upgrade |
| Opus 5.5 / GCP CVP Issue 23 |
No Sep 23–24 enrollment delta — docs still last updated Sep 22 | Watch |
Critical · vLLM ≤0.29.x still exposed
Sep 23 Moderate GHSAs (structured-output EngineCore kill, GLMGA video DoS, Rust Prometheus cardinality, Qwen2/3-VL max_frames) stack on Issue 23’s Sep 22 NIXL/Mooncake batch. Land on ≥ 0.30.0; do not leave multimodal / Rust frontend / structured-output tenants on ≤0.29.x.
Watch · pin v0.5.0 vs tip b11158 · Bedrock connectors
Prefer llama.cpp v0.5.0 for pin-able installs; tip b11158 (Vulkan Adreno) / b11156 (Ling 3.0 VL) are nightlies. Bedrock Salesforce/Zendesk connectors need credentials and incremental-sync review before production RAG cutover.
Opportunity · Omni Cursor ext · Antigravity local · Gemini TTS
Install CloudWatch Omni’s free Cursor/VS Code extension for local agent instrument/eval. Smoke Antigravity LiteRT Gemma 4 26B or Ollama-compatible local backends. Wire Gemini 3.8 Flash TTS / Flash-Lite for voice agents. Optional one-liners: Unity Catalog Pages (Sep 22 companion to Genie One); Amazon Connect Customer A2A (Sep 22).
llama.cpp v0.5.0 Metal MoE / multi-bind on Apple Silicon
Pin v0.5.0 macOS arm64 assets. Release cites Metal MoE + SSM_CONV fusion (#28948). Optional: llama-server --host with comma-separated binds (#28690). Tip-only: b11156+ for Ling 3.0 VL.
Ollama 0.34.4 GA MLX Qwen 3.8 PP on Mac Studio
Leave the RC; install GA. Compare Qwen 3.8 MLX prompt-processing vs 0.34.3. Optional Hub smoke: mlx-community/Qwen3.8-Flash-Next-4bit-mtp — community quant, verify MTP/arch support first.
Issue 24 · Thursday, September 24, 2026 · Not investment advice · Sources cited inline · Other news: none this window · Issue 23 Opus 5.5 / Sol-Luna / Genie One MCP / Switchyard 0.3.0 / oMLX / tip b11126 not re-led
