
Microsoft's BitNet b1.58 Runs a 100B AI Model on a Single CPU
Microsoft's bitnet.cpp runs 100B-parameter models on a single CPU with no GPU, using ternary weights that replace float math with add/subtract ops
20 stories

Microsoft's bitnet.cpp runs 100B-parameter models on a single CPU with no GPU, using ternary weights that replace float math with add/subtract ops

vLLM v0.26.0 ships 411 commits with Day-0 Inkling support, per-group attention backends, tiered KV offloading, and sweeping DeepSeek-V4 speedups across all hardware
Public benchmarks shortlist a model. A small task-specific eval tells you whether it deserves a larger test.

Flash Linear Attention is the open-source kernel library quietly powering production LLMs like Qwen3.5 and Kimi, with 30+ subquadratic architectures in one place

jcode is a Rust-native coding agent harness that boots in 14ms, runs 10+ parallel agents on 8GB of RAM, and lets the agent rewrite its own source code — all free and open-source.

Midjourney makes V8.2 the new default model, promising bolder aesthetics, stronger personalization, and improved image quality over V8.1

APEX quantization compresses Kwaipilot's 35B agentic coding MoE into GGUF files that run locally on consumer GPUs while preserving benchmark accuracy.

Exa's AI-native web search is now a one-command install inside xAI's Grok Build terminal agent, bringing real-time search and deep research to coding workflows

ChatGPT Work agents can now reach login-gated websites via a human takeover step, with sessions that persist across runs

Claude Opus 5 lands in Lovable with a 22% coding boost over Opus 4.7, same price as its predecessor, and near-Fable-5 intelligence at half the cost

Claude Opus 5 launches with a 3x ARC-AGI-3 lead, Fable 5-class intelligence at half the price, and Anthropic's strongest alignment scores yet

A community-quantized build of Poolside's 33B coding MoE hits 1.4M downloads, using a MoE-aware compression scheme that beats Q8 at half the size.

vLLM's new AFD Plugin splits Attention and expert computation into independent services, unlocking +11% decode throughput for large MoE models like DeepSeek-V3.2
Midjourney acquires astrology app Co-Star and names its CEO as Chief Design Officer, signaling a major push into consumer software products

Runway Agent now builds and edits node-based video pipelines from plain text descriptions, removing the need to wire nodes manually.

v0 now takes a single Figma file link and autonomously explores every page and frame to build a complete, multi-screen working app

TileLang is a Python DSL that lets you write GPU kernels with Triton-like ease but CUTLASS-level control — and DeepSeek just used it in production.

Kimi Code CLI 0.29.1 adds per-subagent model routing, OAuth-free web services, and a fix for vLLM reasoning content loss
We tested Gemini 3.6 Flash on 13 debugging tasks and compared its accuracy, cost, speed, and token usage with eight other models.

Sakana AI's Fugu-Ultra v1.1 upgrades its multi-model orchestration system with gains of up to 7.9 points across benchmarks, at the same price as v1.0