Qwen3.8-27B Goes Open Source, Runs on Consumer GPUs
Qwen3.8-27B goes open source: Apache 2.0, 1M context, beats Qwen3.7-Plus on coding and office tasks on a consumer GPU.
Qwen3.8-27B goes open source: Apache 2.0, 1M context, beats Qwen3.7-Plus on coding and office tasks on a consumer GPU.
Google dismantles DeepMind autonomy as teams shift to corporate reporting and Brin returns to AI power.
Gemini 3.7 Flash arrives 3 weeks after 3.6 at half price, pushing FrontierCode to 43.6% and redefining agent economics.
Opus 5 scored 30.2% on ARC-AGI-3, then 96.2% with a computer. Stronger models need simpler harnesses.
AI agent funding crashed in one summer. Here is why, and the B2B and overseas playbooks still generating revenue.
OpenAI Ultrafast runs GPT-5.6 Sol at 750 tokens per second on Cerebras wafer-scale chips, 14x faster with no IQ tradeoff.
ChatGPT Desktop lands on Linux: ChatGPT, Work and Codex in one app. Install the .deb or .rpm and let Codex touch your repo.
Run DeepSeek Harness with one command: an MIT-licensed agent runtime where models, tools and UI are hot-swappable plugins.
DeepSeek Harness is now open source: an everything-is-a-plugin agent framework on the Cordis microkernel with swappable parts.
DeepSeek V4 Pro stable ships, Grok 4.6 undercuts rivals, Claude clears Hadamard matrices, Gemini hits 1B users.
Claude plus three mathematicians built the first order-668 Hadamard matrix, clearing every open case below 2000.
XunZi, an AI biologist trained on 24.4M papers, flags CHK2/IRAK4 as Parkinson targets, confirmed in mouse models.