Released Oct 20 under Apache 2.0, DeepInfer runs DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs or Huawei Ascend 950 NPUs with DeepSeek's own kernels, expert parallelism, speculative decoding, deterministic sampling, and structured outputs. It's offline batch inference only for now: KV-cache management, prefill/decode disaggregation, and load balancing stay internal (The New Stack).
Merged by Linus Torvalds on Oct 21, the docs 7.4 pull adds a top level AGENTS.md symlink to the kernel README so agents find the contribution rules, and removes the Assisted by: attribution requirement and its checkpatch handling, as agreed at the Maintainers Summit. AI agents still must not add Signed off by ; only the human submitter certifies the patch (pull request)…
The Oct 21 release adds /claude api api version to move your Claude API code to the new 2026 10 21 version, plus a claude code.safety event OpenTelemetry event for every tool call a permission rule, auto mode, hook, or sandbox blocked or prompted. It also stops HTTP MCP tool calls being sent twice when a connection drops mid call (only read only or idempotent tools retry now), and Code Review gains @claude review stop to end an always on PR subscription…
On Oct 21, an attacker controlled build pushed litellm==1.107.0 to PyPI (live 11:39 to 13:23 UTC) and to the latest , main latest , and main stable Docker tags, adding a .pth startup hook that steals cloud keys, SSH keys, and /.cursor , /.claude , and /.kiro configs on any Python start. If you installed it, remove cook.pth , cook.py , and sitecustomize.py from site packages, rotate credentials from a clean machine, and pin 1.106.2 (OSV advisory)…
Released Oct 21, requests on the new anthropic version return a 400 if they send temperature , top p , or top k , put document blocks in assistant messages, or attach citations to user text blocks; 2023 06 01 keeps working if you need the old behavior. New SDKs (Python 2.0.0, TypeScript 1.0.0, Go 2.0.0 and others, now Apache 2.0) send it by default (release notes)…
Auto mode now works on Windows 11 24H2+ for personal Google Accounts ( mode auto or Shift+Tab ; Agent stays the default), and it now checks every subcommand in a pipeline and prompts before destructive Git or file commands. MCP tools/call requests also carry meta["antigravity.google/root conversation id"] , so servers see the top level conversation even when a nested subagent makes the call…
Spotted something we missed? Start a thread.