Colibri runs 2.8T-parameter MoE models on consumer hardware

Meta says Muse Wonder models are too large for many people to use and carry potential safety concerns, so they will not be open-sourced. Open weights for Muse Spark are coming soon.

Anthropic migrated its monorepo off git to an in-house version control system called Scion in five months, after daily merges grew from ~2,500 in May 2026 to over 30,000 by September. It also shipped a code review surface called Antique, code search, and a source control IAM system in the same period.

MCP support now ships in the main langchain.mcp package, built on FastMCP for the 2026-07-28 spec, replacing the separate langchain-mcp-adapters install. Elicitation is surfaced as a LangGraph interrupt and tool lists are now cacheable, matching the spec's stateless core.

LangChain's Deep Agents evals found GLM-5 (z.ai) and MiniMax M2.7 score similarly to closed frontier models on file operations, tool use, and instruction following. Outputting 10M tokens/day costs ~$250/day on Opus 4.6 versus ~$12/day on MiniMax M2.7 — about $87k annually.

LangChain automated migrating inline docs code samples into standalone, CI-tested files using its Deep Agents CLI, so broken examples fail the build. The manual path otherwise requires extracting snippets, adding setup/teardown, snippet markup, and CI wiring.

Jensen Huang said at a Paris event that Nvidia will announce $2 billion in European investments over the next week, on a tour that also takes him to London and Berlin.