Cloudflare Infire — disaggregated LLM inference beats vLLM by 20%, Unweight cuts model size 22%
Cloudflare Infire (Rust) uses disaggregated prefill/decode to beat vLLM 0.10 by 20% on H100s. Unweight achieves 15–22% lossless model weight compression.
Tags · #release
Google Jules public beta — async coding agent goes global with CLI and GitHub Action
Jules (Gemini 3 Pro) is now global public beta with a GitHub label Action and Jules Tools CLI — the first async GitHub-native coding agent rival to Claude Code.
Microsoft Agent 365 GA + Agent Framework 1.0 — enterprise agent control plane is live
Agent 365 GA at $15/user: per-agent Entra IDs, Defender MCP blocking. Agent Framework 1.0 is the open-source multi-agent baseline with A2A and MCP interop.
NVIDIA GR00T N1.7 goes commercial — robot foundation model leaves research-only status
GR00T N1.7 VLA moves to commercial early access, removing research-only limits. Jensen Huang previewed GR00T N2 claiming 2× task success vs. current top VLAs.
Copilot in Visual Studio: cloud agents, Debugger agent, .claude/skills/ integration
Copilot in VS gains cloud agent sessions, profile-level custom agents, .claude/skills/ loading, and a Debugger agent that reproduces issues at runtime.
Cursor SDK enters public beta — agent harness as shared infrastructure
Cursor released a TypeScript SDK giving programmatic access to the same runtime, harness, and models powering its desktop, CLI, and web apps.
Mistral releases Medium 3.5 with 256K context, 77.6% SWE-Bench Verified
Mistral Medium 3.5: 128B dense, 256K context, 77.6% SWE-Bench Verified. Vibe gets cloud remote agents; Le Chat gets a Work Mode.
vLLM v0.20.0 ships DeepSeek V4 + PyTorch 2.11 + FlashAttention 4
vLLM v0.20.0: 752 commits, 320 contributors. CUDA 13, PyTorch 2.11, Transformers v5, Python 3.14, FlashAttention 4 default, 2-bit KV cache.
Cursor 3.2 ships /multitask, worktrees, and multi-root workspaces
Cursor 3.2 introduces a /multitask command that spawns parallel async subagents, expands worktrees in the Agents Window, and adds multi-root workspaces.
OpenAI launches GPT-5.5 (codename Spud) with 1M context, $5/$30 pricing
GPT-5.5 (codename Spud) shipped to ChatGPT and Codex Apr 23, API Apr 24. OpenAI cites Terminal-Bench 2.0 82.7% and FrontierMath gains over Opus 4.7.
Anthropic launches Claude Design — codebase-aware visual prototyping
Claude Design (research preview) generates prototypes, slides, and one-pagers from natural language, and reads codebases to apply design systems.
Anthropic releases Claude Opus 4.7 with improved coding and 3.75 MP vision
Claude Opus 4.7 ships to Claude products, API, Bedrock, Vertex, and Microsoft Foundry. Better coding, ~3x higher vision resolution, same pricing.