Dev Stack
PyTorch 2.15 speeds MoE decoding on Macs up to 6.5x
PyTorch 2.15 adds FP8 tensors on Apple GPUs and new Metal kernels. Qwen3.6-35B-A3B decoding on an M3 Ultra jumps from 9.4 to 61.3 tokens a second.
Local LLM news and analysis
Running language models locally instead of in the cloud: Ollama proxies that put an open model behind a desktop app, hardware judged by memory and bandwidth, Windows modes for local models, and on-device search products. Each story states the memory needed and what works.
AI & LLMs
AMD GAIA 0.25 folds its local AI agents into one
AMD's open-source local agent app now runs one agent, shrinks its Windows Python setup from 0.9GB to 0.1GB and stops crashing on 16GB laptops.
New & Next
Nvidia RTX Spark laptops go on sale from $2,599
Nvidia RTX Spark laptops from six makers went on sale on October 16 from $2,599, with up to 128GB of memory and a new 74B open model, Nemotron 3.5 Bolt.
Dev Stack
Intel LLM-Scaler-vLLM 1.5 jumps to vLLM 0.26 on Arc GPUs
Intel's LLM-Scaler-vLLM 1.5 moves its Arc GPU Docker stack from vLLM 0.14 to 0.26 and adds Gemma 4 and Qwen3.8 models plus FP8 and INT4 options.
New & Next
Surface Laptop Ultra with RTX Spark starts at $2,599
Microsoft's Surface Laptop Ultra pairs Nvidia's RTX Spark chip with up to 128 GB of shared memory. It costs from $2,599.99, with preorders open and shipping October 16.
Dev Stack
Surface RTX Spark Dev Box opens for preorder at $5,999
Microsoft opened US preorders for the Surface RTX Spark Dev Box at $5,999 on October 7, 2026, with 128GB of unified memory for local AI models.
AI & LLMs
Google AI Edge Foresight takes meeting notes offline
Google AI Edge Foresight, an experimental Mac app, turns shorthand into meeting notes on-device with Gemma 4 and the 740M-parameter EmbeddingGemma 2.
AI & LLMs
EmbeddingGemma 2 puts multimodal search on a phone
Google DeepMind's EmbeddingGemma 2 maps text, code, images, video and audio into one vector space with 740M parameters. Text-only use needs about 191 MB of RAM.
AI & LLMs
Strands Decider 2B makes agent choices in 115 ms
AWS's Strands Decider 2B is an open-source model that picks from a list instead of writing text. It answers in 115 ms on an RTX 3090 and scores 72.3% on JevBench.
Dev Stack
RemoveMacAI strips Apple Intelligence models from macOS 27
A new MIT-licensed command-line tool turns off Apple Intelligence on macOS 27, deletes its on-device models and stops macOS from downloading them again.
AI & LLMs
Aleph Alpha Kolibri 1: open German-English MoE model
Aleph Alpha released Kolibri 1 on October 3: a 78.1B-parameter open model with 3.46B active, built for German and English, under the Apache 2.0 license.
New & Next
Framework Desktop 192GB pre-orders open at $6,799
Framework opened pre-orders on September 30, 2026 for a 192GB Framework Desktop with AMD's Ryzen AI Max+ PRO 495, at $6,799 and shipping in November.
AI & LLMs
Liquid AI LFM2.5-VL-DSpark speeds up its 3B vision model
Liquid AI's 280M-parameter LFM2.5-VL-DSpark draft model decodes its 3B vision model up to 3.13x faster on an M5 Max, with identical output.
AI & LLMs
Qwen3.5-9B spends 32KB per token on KV cache
Qwen3.5-9B runs full attention on only 8 of its 32 layers, so its KV cache costs 32KB per token. That ratio decides whether it fits in 8GB.
Dev Stack
Intel's Linux NPU driver 1.38 adds Ubuntu 26.04 support
Intel's NPU driver 1.38 officially supports Ubuntu 26.04 LTS about five months late, with 322 tests passing across five processor generations.
Dev Stack
Microsoft's Project Zenith is a Windows mode built for local AI coding
Project Zenith is a developer-focused Windows 11 experience for running 30B-plus parameter models locally, on PCs with 64GB of unified memory.
Dev Stack
New Mac mini and Mac Studio put more memory bandwidth behind local AI
Apple announced new Mac mini and Mac Studio models on August 25, 2026, shipping September 22. For local model work, the memory ceiling matters more than the AI multipliers.
Startup Watch
Clipto raises $15M at a $250M valuation for on-device AI search
Clipto said on August 31, 2026 that it raised $15 million at a $250 million valuation. Its search index runs on the user's own machine instead of the cloud.
Dev Stack
Ollama 0.33 lets Claude Desktop run Qwen, DeepSeek and Kimi locally
Ollama 0.33.0 adds a local proxy that swaps the model behind Anthropic's Claude Desktop app for an open one. A spring attempt failed on model-ID checks.
The weekly digest
One email every Friday with the week's top stories from all six desks: AI, dev tools, coding, the tech industry, startups and what's next. Free, no spam.