Fetching from the wire…
01
02
03
04
05
06
07
08
Infra2026-09-24 · source-backed
The only v-prefixed tag in its last 100 releases, with structured notes splitting API changes, new models, server changes and ggml changes (GitHub). Highlights: CUDA conv2d via implicit GEMM, Metal MoE and SSM_CONV fusion, input_image in server function-call outputs, HRM-Text/DFM Mimir 1B support, ggml 0.25.0 with RPC protocol v7. Downstream packagers finally get something to pin to. Build b11158 also enables KHR cooperative matrix for Qualcomm Adreno GPUs, which lines up with the Snapdragon news above.
Each link below shares sources, entities, or timing with this story.
Build b10687 defaults the xmem F16xF32 GEMM on for Adreno X2E after measuring kernel_mul_mm_f16_f32_l4_lm running gpt-oss-20b attention projections at about a quarter the speed of the tuned dense q4_0 GEMM. Worth 25% prefill. The route already existed but was opt-in, so nobody...
Build b10758 (September 2) fuses matmuls landing on Hexagon HMX, fuses MUL_MAT_ID into MUL_MAT_ID_NX, and adds VA defragmentation so large-dim runs abort less on fragmented address space (release). Build b10757 handles batch sizes above 4 for IQ3_S mat-vec when NUM_COLS > 4, r...
Cross-platform desktop STT built with Tauri. Fully offline using Whisper and Parakeet models. GPU-accelerated on CUDA, or CPU-only via Parakeet V3. Designed to be *"the most forkable speech-to-text app."* Competes with paid tools like Wispr Flow with zero cost and full privacy...
Build b10677 fixes ggml_vk_graph_optimize, where is_src_of didn't treat two views of one tensor as dependent, so the optimizer reordered nodes across aliased reads and writes. Maintainers describe the result as silently wrong tokens under greedy decoding, different output on e...
1. Build a Private Claude Code Plugin Marketplace (intermediate, vibe-coding) — Bundle skills, agents, hooks, MCP servers into installable team plugins via GitHub repos. Docs 2. Google ADK TypeScript Multi-Agent Orchestration (intermediate, agent-patterns) — Code-first agent f...
Build b10715 folds the encoder into the decoder's embd branch so one llama_decode handles it. The encoder is a fully-connected layer plus a norm, but running it as a separate llama_encode forced a round trip before the injection decode could re-upload it, and built a second gr...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.