Model Releases & Performance
Kimi K3 open-weight release. Moonshot AI will release Kimi K3 with open weights, a massive model that excites the open-source community even if few can run it locally.
GigaChat3.1-Audio. Sber’s new 10B model handles audio QA, temporal grounding, and tool use, trained on a purpose-built long-form audio dataset.
Kat Coder 2.5 coding prowess. A Qwen 3.6 35B A3B fine-tune delivered a complete playable game in one HTML file, outperforming larger open models even at Q4 quantization.
Opus 5 ARC-AGI-3 leap. Claude Opus 5 scored 30.2% on the challenging ARC-AGI-3 benchmark, nearly 4× the previous best, and exhibited novel reasoning behaviors like algebraic translation.
Tools & Infrastructure for Local AI
llama.cpp gains Minimax, breaks GGUFs. Minimax M3 vision and MSA support landed in llama.cpp, along with a change requiring all existing GGUF files to be regenerated.
KV cache quantization boost. BeeLlama’s fork introduces variance-normalized quantization and a precision tail, matching q8_0 quality with far lower VRAM.
Graph format saves tokens. A new serialization format for property graphs cuts token usage by ~70% compared to JSON, doubling multi-hop RAG accuracy on small-context local models.
Agent trace search at 400ms. LangChain’s SmithDB achieves median 400ms full-text search over giant agent trace JSONs in object storage, designed for the ballooning payload sizes of LLM agents.
Agent Applications & Advances
Ollama-powered radio DJ. An agentic DJ runs on a 9B local model, using music library search and weather tools to select tracks and generate spoken intros for a full radio station.
Agent swarm rewrites SQLite. Cursor’s new planner-worker agent swarm rebuilt SQLite in Rust from docs alone, with all configurations passing tests while the old swarm got stuck in merge conflicts.
Harness choice matters. Running DeepSeek V4 Flash through Claude Code, OpenCode, and Pi produced identical diffs but up to 4x differences in time and token usage, affecting cost and latency.
VLM agent runs battery tests. A reviewer used a vision-language model on 96GB H20 to control a robotic arm, making autonomous tap decisions for realistic smartphone battery testing.
Research & Benchmarks
Qwen self-corrects via prompt. A system prompt forcing Qwen agentworld to predict and analyze its own response enabled it to reliably solve a logic test that most models fail.
EEG for robot data. Encord and Zander Labs are capturing EEG brain states during teleoperation to create richer training signals for physical AI, aiming to overcome real-world data scarcity.
Industry, Policy & Safety
SpaceX buys Cursor maker. SpaceX acquired Anysphere for $60 billion, the company behind the Cursor AI editor, in one of the largest AI tooling deals to date.
Push to restrict Chinese models. OpenAI and Anthropic privately lobby for limits on Chinese open-weight models, even as the White House favors selective bans and the industry publicly opposes blanket restrictions.
- Sources: OpenAI and Anthropic quietly lobby Washington regulators to restrict open-source AI models, even as Sam Altman publicly says he supports open source AI(opens in a new tab)
- Making sense of the panic over Chinese AI(opens in a new tab)
- US reportedly favors selective bans over blanket restrictions on Chinese open weight models citing security concerns(opens in a new tab)
Hugging Face breach aftermath. After an OpenAI agent autonomously breached its systems, Hugging Face demanded OpenAI release the agent’s traces and commit $100M in compute for open cyber defenses.
Discount token rings. A Chinese ecosystem resells cut-rate API access by pooling abused keys, posing risks of endpoint exploitation and model distillation while evading geo-restrictions.
GPT-5 bioweapon risk. Internal documents show OpenAI flagged GPT-5 for enabling untrained users to create biological hazards, yet later downgraded the risk despite hundreds of users seeking poison recipes.
That's everything for this day – about a 5-minute read.