Microsoft's Frontier Tuning Beats GPT-5.5 at 10x Lower Cost
Microsoft's Frontier Tuning uses private reinforcement learning environments to build custom MAI models that outperform GPT-5.5 at 10x lower cost
20 stories
Microsoft's Frontier Tuning uses private reinforcement learning environments to build custom MAI models that outperform GPT-5.5 at 10x lower cost
Nous Research adds a guided Profile Builder to Hermes Agent's web dashboard, letting you configure identity, model, skills, and MCP servers in one visual flow instead of editing YAML
Nathan Lambert, the researcher behind Ai2's OLMo and Tulu and author of the RLHF Book, joins Arcee as Research Advisor as the open-source AI movement faces its most critical inflection point yet.

gws is a Rust-built, open-source CLI that unifies every Google Workspace API behind one command surface, with 100+ built-in AI agent skills and dynamic command generation from Google's Discovery Service.
Replit's new Package Firewall, built with Socket, blocks malicious npm and pip installs at the network level -- stopping 8,000 attacks per day before a single line of malicious code runs.

Google's DiffusionGemma ditches token-by-token generation for parallel diffusion, hitting 1,000+ tokens/sec on a single H100 -- at the cost of benchmark quality

Google's DiffusionGemma generates entire 256-token blocks at once, hitting 1,288 tokens/sec on H200 — but trades benchmark accuracy for raw speed
Dynamic workflows in Claude Code go GA, while scheduled deployments and credential vaults hit public beta in Claude Managed Agents

Google DeepMind's DiffusionGemma generates 256 tokens in parallel, hitting 1,000+ tokens/sec on H100 — but trades raw quality for speed
Google's DiffusionGemma generates 256 tokens in parallel using text diffusion, hitting 1,000+ tokens/sec locally — at the cost of some output quality
Reka absorbs Moonvalley's 50-person team of ex-DeepMind researchers to build WLAM, a world model that lets robots simulate before they act

PaddleOCR's new PP-OCRv6 beats GPT-5.5, Gemini, and Qwen3 on text detection with just 34.5M parameters, while running 5x faster on CPU
Kyutai uses RL post-training to teach Moshi and PersonaPlex when to speak, stay quiet, and say 'uh-huh'

Kimi Code CLI v0.13.0 adds custom color themes, one-command Claude Code and Codex migration, and live plugin update alerts

A single Python import wraps ten face recognition models plus twenty detectors, hitting human-level accuracy without any pipeline plumbing.

NVIDIA's CUDALibrarySamples repo now includes runnable examples for cuPQC, cuEST, and nvMatmulHeuristics — three recently released CUDA-X libraries covering post-quantum crypto, quantum chemistry, and GEMM auto-tuning.
Alibaba's Wan platform adds a one-click Fisheye Lens effect to its growing suite of AI image editing tools, no hardware required.

Tencent open-sources WorldPlay, the first real-time interactive world model that generates explorable 3D environments at 24 FPS with long-term geometric consistency
Vercel Labs open-sources Zero, a graph-first systems language where the compiler speaks JSON so AI agents can write, check, and repair code without human translation
Harness is a free Claude Code plugin that turns a plain-English description into a full multi-agent team, complete with agent definitions, skills, and orchestration logic.