Re-measured the quality and long-context speed of Qwen 3.6-35B-A3B KV cache quantization on M5 Max.
Built a compiler that lowers PyTorch graphs down to CUDA using 5,000 lines of Python and raw CUDA.
Starting with Ubuntu 26.04, Canonical will let users install CUDA and ROCm with just apt.
HyperResearch, an open-source tool that extends Claude Code into a high-performance deep research framework, has been released.
A TurboQuant inference path was added to Runtime Valkyr, achieving 120 tok/s on Gemma with an RTX 3090.
An Nvidia vice president said AI costs now exceed labor costs.
Chirp has been released as an offline TTS app supporting Kokoro and Qwen3-TTS.
LLM 0.32a0 has been released with a message- and stream-based redesign that maintains compatibility.
Version 0.32a0 of the command-line LLM tool llm has been released.
Opening arguments have begun in Musk's lawsuit against OpenAI and Microsoft.
It's a header-only C library that unifies multiple LLM providers including OpenAI and Anthropic.
Claude Code with MCP code health guidance increased refactoring improvements by 2-5x.
OpenAI unveiled a five-pillar action plan for AI-powered cyber defense.
Sensitive AI training data proved vulnerable to membership inference and reconstruction attacks, with DP and MPC serving as defenses.
With a 2x RTX 3090 Ti setup, applying `-sm tensor` increased llama-bench throughput.
AssemblyAI has launched an API for voice agents.
Warp has open-sourced its source under AGPL and transitioned into an open agentic development environment.
inclusionAI has open-sourced Ling-2.6-1T, a 1T-parameter-class model.
OpenAI will hold DevDay 2026 in San Francisco on September 29.
IBM Granite has released Granite Speech 4.1, an enterprise-grade multilingual transcription and translation model.
Warp has open-sourced its client under AGPL and welcomed OpenAI as founding sponsor.
IBM Granite 4.1 was released under Apache 2.0, trained on 15T tokens and extended to a 512K context.
Static benchmarks keep getting compressed, but agent and training evals are seeing costs skyrocket.
LangChain has added model-specific profiles to Deep Agent, improving benchmark performance by 10-20 points.