Georgi Gerganov, creator of llama.cpp and lead at ggml-org, reports daily use of Qwen3.6-27B for coding tasks over the past month and a half. He runs the model locally on an M2 Ultra Mac and an RTX 5090 box, using a stripped-down pi agent harness with the -nc --offline flags and a short system prompt. He finds it helpful for small maintainer tasks and affirms the model's capability for real-world, local AI-assisted programming. The endorsement comes from a Hacker News comment on the article "Running local models is good now" by Boykis.
The 0.3a0 release of datasette-agent introduces the execute_write_sql tool, which writes to a database after requesting user approval and respecting user permissions. The chat terminal mode now supports user approvals and three new options: --root, --yes, and --unsafe for auto-approval. Tools can provide plain text alternatives to HTML for CLI display. Users can now directly chat with a specific database and modify it via prompts like 'create a notes table' using the --unsafe flag.
The Datasette 1.0a33 alpha release extends the existing ?_extra= URL parameter pattern, previously only available for tables, to also work with SQL queries and individual rows. This new API behavior is now fully documented. Simon Willison built a custom API explorer tool to demonstrate the feature, using Claude Fable 5 for planning and GPT-5.5 xhigh for implementation. The release represents a significant step towards the stable 1.0 version of Datasette.
Version 0.2a0 of datasette-agent is released. Tools can now ask user questions mid-execution using await context.ask_user(...), supporting yes/no, multiple-choice, or free-text input. The question renders as a form in the chat UI and persists to the internal database, so suspended conversations survive a server restart; the tool re-executes from the top once answered. A built-in save_query tool saves SQL as a Datasette stored query, but requires explicit human approval with full SQL, name, database, and visibility displayed before storage. This release was built on a new LLM alpha developed with Claude Fable 5.
Google has released the diffusiongemma-26B-A4B-it model under an Apache 2.0 license, building on its earlier experimental Gemini Diffusion. It is openly available on Hugging Face and NVIDIA offers free access via their NIM cloud API, demonstrating over 500 tokens per second generation speed. In a test, the model generated 2,409 tokens in 4.4 seconds, highlighting its efficiency for text generation tasks.
Version 0.32a3 of the open-source LLM command-line tool llm has been released. The update was almost entirely written by Anthropic's new Claude Fable 5 model. Developer Simon Willison detailed the experience in a separate write-up, highlighting the model's code generation capabilities.