Fetching from the wire…
OSS2026-08-11 · source-backed
h3.c, published August 9 by the Redis creator, targets M3/M5 Max Macs from scratch. Reported: 512×512, 22-frame, 20-step generation at ~16.69s on M5 Max, dropping to ~12.60s with token reduction, 8.02s for a native 256-square preview, 19.32s on the int8 path. Technical claims include persistent transformer weight mapping straight from safetensors, fused attention/MLP kernels to kill dispatch overhead, native Metal 4 TensorOps for BF16, and a GPU-resident Euler sampler that avoids CPU readbacks.
Each link below shares sources, entities, or timing with this story.
9,001 stars and 1,258 forks, a 14% fork ratio, Swift under MIT, with 1.0.12 tagged August 29. It repurposes Apple's Virtualization.framework PV=3 paravirtualization mode, introduced in macOS 15, to expose SEP, synthetic battery, USB touch and PL011 serial so a stock iOS kernel...
ANE is trending at 7,070 stars for doing what Apple doesn't permit: training, not just inference, on the ANE. Core ML exposes it for inference only, so every on-device fine-tune to date has fallen back to GPU via Metal. If it holds up, the power and thermal math for local trai...
DeepSeek-V4-Flash-0731 landed July 31 under MIT with a DSpark speculative-decoding module attached. Terminal Bench 2.1: 82.7. Toolathlon-Verified: 70.3. DSBench-FullStack: 68.7. DeepSWE: 54.4. NL2Repo: 54.2. The model card claims it beats DeepSeek-V4-Pro (Preview) "despite its...
milind-soni/OpenMausBot (724 stars since August 11, Electron + React 19, MIT) makes each sidebar bot a real local Claude or Codex agent with its own personality, model, cloud computer, and connected apps, with approval gates. Distributes as a signed notarized one-click .dmg fo...
Deep reverse engineering of Apple's M4 Neural Engine revealing undocumented tiling strategy, memory bandwidth constraints, and why certain architectures run faster on ANE vs GPU. Essential for on-device AI inference optimization. HN
For about a year, "run your agent locally" meant accepting a model that couldn't reliably call a tool twice in a row. That excuse is gone. Meta Superintelligence Labs published Muse Glimmer today: a 29.6B dense causal transformer, 52 layers, 6,656 hidden dim, with a ~1.8B ViT-...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.