Articles
Why first — then the essentials of how (stack, quantization, MCP).
TypeSafe ships Jev, a System One model: typed decisions, no text. LangChain turns it into a harness classifier — routing and auto-mode — at a fraction of chat cost.
PrismML recompresses Qwen3.8-27B into true ternary weights (1.72 bits). 5.95 GB, 98.2% of FP16, ~47 tok/s on an M5 Max. The catch: it will not run in stock llama.cpp.
Brooklyn (Nous Research) drops the agent as an overlay on a WoW session. The HUD is no longer a window you alt-tab to: it’s a layer, and its position is the context.
Inco AI’s parallel drafter for Qwen3.8-27B: +20% acceptance vs DFlash, 2.7–3.4× autoregressive throughput, identical output. Why it matters as soon as an agent loops.
Qwen3.8-27B specs, official scores versus 3.6 and Opus 4.6 Max, and what that changes for local AI.
Studio, Desktop, and Unsloth kernels: GGUF / MLX, low-VRAM fine-tunes, Data Recipes — and why the Qwen3.8-27B GGUF fits in 17 GB.
A Next.js site that turns public Hub activity into an interactive Pokémon TCG card — holo shaders, follower binder, PNG export, and a shareable profile page.
Notes from the Hugging Face live with Daniel Hanchen (UnslothAI): Studio, dynamic quants, benchmarks, and low-VRAM fine-tuning on Mac/Windows/Linux.
Fable 5 just landed and tops the Cursor benchmark at a steep price: I’m shifting agentic work to Hermes and affordable models.
Free tokens instead of Cursor models — Qwen 3.6 27B on AWS via vLLM (4-bit), Ollama locally for git-mentor.
Give the agent concrete GitHub actions instead of one big “look at my repo” prompt.
Built in a hurry when MCP landed — so Cursor could call the real Red Bee Media APIs I need at work.