SGLang v0.5.17 ships Kimi K3 day-0 + Weaviate fixes, Anthropic adds session budgets

🚀 Models & Releases

SGLang v0.5.17 ships day-0 support for Kimi K3, a 2.8T-parameter multimodal LatentMoE (896 experts, top-16 routing in 3584-dim latent space) with 1M-token context, 69 KDA linear-attention layers, 24 MLA layers, and MoonViT3d vision tower. Native MXFP4 checkpoint available. (sgl-project/sglang#v0.5.17)

Weaviate v1.38.9 patches backup chunk integrity, usage module, and async replication. No breaking changes, but fixes may impact production stability. (weaviate/weaviate#v1.38.9)
Anthropic API now supports session budgets for Claude Managed Agents: set a hard spend cap (priced at public rates), pause sessions with budget_reached stop reason, and resume by adjusting the budget. Applies to deployments per session. (Anthropic API docs)

llama.cpp b10326: TTS timing now accounts for vocoder pass, ensuring accurate audio-to-process ratios. (ggml-org/llama.cpp#b10326)
OpenAI Codex 0.147.0: Adds portable Agent Plugins, cross-catalog search (local/personal/workspace/remote), and persistent conversation sections with incremental transcript browsing. (openai/codex#0.147.0)
cline Desktop v0.0.10: Remote MCP servers now support OAuth authentication (Settings → MCP), including pre-registered clients and token invalidation on config changes. (cline/cline#v0.0.10)
Anthropic SDK Python v0.121.0: Adds mid-conversation-tool-changes-2026-07-01 beta and support for Claude’s new session budget feature. (anthropics/anthropic-sdk-python#v0.121.0)
pydantic-ai v1.107.2: Security patch enforces 50 MiB cap on remote content downloads via web_fetch or FileUrl to prevent memory exhaustion. (pydantic/pydantic-ai#v1.107.2)

Selective Context Preference Optimization: New method trains models to ignore misleading external signals while retaining context when trustworthy. Addresses failure modes in RAG and tool-use systems. (arXiv:2608.06377v1)
The Bitter Lesson of Tool Calling: Evaluates programmatic tool calling (vs. rigid JSON) for LLMs, highlighting trade-offs in flexibility and reliability. (arXiv:2608.06370v1)

OpenAI paused development of its Astra model after it reached a "critical cybersecurity threshold"—capable of autonomously executing attacks on hardened systems. (TechCrunch)
Rippling launched AI Spend Console to track employee AI usage and ROI after internal spending surged. (TechCrunch)
Cloudflare released Kitesurf, a cloud-hosted browser optimized for AI agents (lower compute than Chromium for automation tasks). (TechCrunch)

Hank Green (YouTube creator) announced a personal AI policy, urging peers to adopt transparency standards. (Business Insider)
Students proposed a National AI Policy Framework covering education, ethics, and governance. (Education Week)
Texas politicians unveiled competing AI regulation plans for Congress, focusing on federal oversight and industry guardrails. (KXAN Austin)

SGLang v0.5.17’s Kimi K3 support is a production-ready drop-in for massive multimodal workloads—benchmark if latency/throughput aligns with your stack.
Anthropic’s session budgets let you enforce hard spend caps on agentic workflows, critical for cost-sensitive deployments.
Weaviate v1.38.9 fixes are mandatory for teams relying on async replication or backup integrity in production.