
Thinking Machines Lab's Inkling Hits Databricks as a 975B Open-Weights Rival
Thinking Machines Lab's first open-weights model lands on Databricks with enterprise governance, coding agent integrations, and a 975B-parameter MoE architecture
20 stories

Thinking Machines Lab's first open-weights model lands on Databricks with enterprise governance, coding agent integrations, and a 975B-parameter MoE architecture

Google's always-on AI agent Gemini Spark expands internationally with four new productivity upgrades including Google Docs editing and parallel source retrieval

RadixArk open-sources Miles, an enterprise RL framework built on slime that tackles MoE training collapse, FP8 stability, and rollout throughput at frontier scale.

Thinking Machines releases Inkling, a 975B open-weights MoE model with native audio, vision, and controllable reasoning effort

Anthropic's new study catches frontier AI agents sabotaging code, assisting fraud, and manipulating training labels — in controlled simulations across 14 models

Google patches Gemma 4 with faster vision, FA4 throughput gains, and community-driven fixes for tool calling and lazy responses

OpenAI's GPT-Red uses adversarial self-play to automatically find and fix prompt injection vulnerabilities, making GPT-5.6 Sol 6x more robust than its predecessor.
A voluntary 30-day gate already delayed GPT-5.6 by 12 days. A mandatory version is the next fight.

PyTorch-Triton 3.7 ships a plugin system that loads custom compiler passes at runtime, ending the era of maintaining Triton forks for hardware-specific optimizations

Perplexity's new SPACE platform runs agent sessions in disposable Firecracker microVMs with rolling snapshots, cutting sandbox creation latency 3-5x

Arena.ai launches a factuality-weighted leaderboard built on 2 million labeled LLM claims, reshuffling which models actually sit on top

Anaconda acquires open-source AI coding agent Kilo Code, bringing 3M developers and 10 trillion monthly tokens under one governed, model-agnostic platform

Hume AI's new benchmark evaluates 40+ voice models on human-quality dimensions that word error rate and latency completely miss

Android Studio Quail 2 hits stable with parallel Gemini agent tabs, native LeakCanary profiling, and AI-powered crash remediation

LeetCUDA is an open-source CUDA learning repo with 200+ kernels, PyTorch bindings, and curated GPU programming blogs that has grown to 11.5k stars

vLLM's new connector interface lets TileRT's latency-optimized engine plug in as a swappable decode backend, hitting 618 tok/s single-user on B200s with zero changes to vLLM.

Microsoft open-sources a VS Code extension that reads your AI coding session logs and coaches you on how to use agents better — all locally, zero telemetry.

AI at Meta entered an unnamed model in the Asian Physics Olympiad's theoretical exam and hit a perfect 30/30, tying with the top three human students.

A free, code-first GitHub repo with 28 runnable tutorials covers every layer of shipping a GenAI agent to production, from memory to security to GPU deployment.

Cactus Compute distilled Gemini into a 26M-parameter model that runs function calling on phones and watches at 1,200 tokens/sec, beating models 10x its size