AI News — July 10, 2026

$ cat top-story.md

▸ Meta Ships Muse Spark 1.1 — and Ends Its Free-Weights Era

Meta Superintelligence Labs' upgraded multimodal reasoning model adds a 1M-token context window, parallel subagent delegation, and stronger computer-use and coding performance — but for the first time Meta is charging for it, via a new paid Meta Model API instead of open weights. Why it matters: A real strategic reversal for self-hosters who've relied on free Llama-family weights. // Meta AI / TechCrunch

▸ MiniMax Reportedly Plans a 2.7-Trillion-Parameter Open-Weight Model

Per The Information, Chinese lab MiniMax is preparing "M3 Pro," a 2.7T-parameter model it intends to open-source — over 6x the size of its current 428B MiniMax M3 and larger than any Chinese open model on the market today. Why it matters: If it ships as planned, it resets the ceiling for "download and self-host"; treat as an unconfirmed plan for now, not a release. // The Information (via TheNextWeb)

▸ Graphify — Turn Any Codebase Into a Local Knowledge Graph for Coding Agents

Graphify uses local tree-sitter AST parsing (no LLM calls) to turn code, SQL schemas, docs, papers, images and video into one queryable knowledge graph, plugging in as a skill for Claude Code, Cursor, Codex, Gemini CLI and 15+ other agents. YC S26-backed, just shipped v0.9.11, sitting at 81.4k stars. Why it matters: Gives agents compact, navigable structure over a codebase instead of dumping raw files into context. // GitHub Trending

▸ Colibri — Run a 744B-Parameter Model on 25GB of RAM

A ~1,300-line, zero-dependency C inference engine keeps GLM-5.2's dense components resident in RAM (int4, ~9.9GB) while streaming its 21,504 routed experts from disk on demand — enough to answer, slowly, on hardware costing less than one H100's cooling fan. Why it matters: A genuine proof that expert-streaming can put frontier-scale MoE models within reach of consumer hardware. // Hacker News (Show HN)

▸ SkillOpt — Microsoft's Trainer for Reusable Agent Skills

SkillOpt treats a skill document — not model weights — as the trainable artifact, applying epochs/batches/validation gates to bounded text edits and producing a compact best_skill.md that works with any frozen backend (OpenAI, Azure, Claude, Qwen, MiniMax). v0.2.0 adds "SkillOpt-Sleep," an offline nightly self-evolution loop. Why it matters: A disciplined, measurable alternative to hand-tweaking prompts and skills. // GitHub / Microsoft Research

▸ Claude Code v2.1.205 & v2.1.206: MCP Timeout Fix, New Auto-Mode Safety Rule

v2.1.205 blocks tampering with session transcript files in auto mode and makes it ask before rm -rf on unresolved variables; v2.1.206 fixes MCP servers ignoring per-server request_timeout_ms (which caused long tool calls to time out at 60s) and adds a /doctor check that flags CLAUDE.md content Claude could already derive from the codebase. Why it matters: Fixes two things heavy MCP/skill users hit constantly — premature timeouts and bloated CLAUDE.md files. // Claude Code changelog

▸ Sophos: AI Coding Agents Are Tripping Endpoint-Security Rules Built for Attackers

A month of Sophos telemetry shows Claude Code, Cursor and Codex regularly triggering attacker-detection rules — decrypting saved browser credentials via Windows DPAPI, enumerating Credential Manager with cmdkey, fetching installers via certutil/bitsadmin, and (for Cursor) dropping startup-folder persistence scripts. Why it matters: Explains why your EDR/AV may flag legitimate agent sessions, and argues for scoping detection to the agent's process rather than disabling it. // The Hacker News / Sophos

▸ Abralo — Free Native App to Run Multiple Claude Code Agents in One Window

A lightweight (few-MB) desktop app that drives your own Claude Pro/Max Claude Code sessions, showing one agent per project folder in a single window with token-burn-rate tracking and a nudge when an agent needs input — up to 4 agents free, native builds for macOS/Windows/Linux. Why it matters: A simpler alternative to juggling terminal tabs when running several agents in parallel. // Hacker News (Show HN)

▸ LMCache — A Persistent KV-Cache Layer for vLLM That Speeds Up RAG Workloads

LMCache turns vLLM's KV cache into reusable, tiered, engine-independent storage (CPU RAM, disk, Redis, S3) with non-prefix reuse, claiming up to 3–15x throughput gains on multi-round QA/RAG workloads; v0.5.1 (July 6) added multi-process mode and multi-node P2P CPU memory sharing. Why it matters: A drop-in layer to cut GPU cycles on repeated or overlapping RAG queries against your Gemma deployment. // GitHub / LMCache

▸ Open WebUI v0.10.2: Folder-Structure Knowledge Uploads, SSRF Hardening

Knowledge-base folder uploads now preserve subfolder structure instead of flattening it, reasoning tokens stream live, and outbound HTTP requests block 3xx redirects by default for SSRF protection (configurable via AIOHTTP_CLIENT_ALLOW_REDIRECTS). Why it matters: Changes how documents land for RAG ingestion and adds a security default worth checking before your next upgrade. // Open WebUI GitHub

▸ New Paper: Pooling-Aware Fine-Tuning Shrinks Multi-Vector Retrieval by 83% With No Accuracy Loss

A fine-tuning recipe for late-interaction (ColBERT-style) retrieval models makes a single model work well across multiple compression levels, reporting up to 83% vector compression on BEIR SciFact with no drop in retrieval accuracy versus the unpooled baseline. Why it matters: Directly applicable if you're running or considering multi-vector retrieval alongside dense embeddings in Qdrant — same recall, far less storage. // arXiv

▸ GPT-5.6 SOL is HERE — Matthew Berman

After two months and 25B+ tokens of internal use, Berman reviews GPT-5.6 Sol against Claude Fable 5 on real coding-agent work, benchmarks and pricing ($5/$30 per Mtok vs. Fable's $10/$50). Why it matters: An independent, hands-on comparison beyond the press-release numbers. // YouTube (Matthew Berman)

▸ AI #176 Part 1: Doing It Live

Zvi Mowshowitz's weekly roundup covers the new model gap opened by GPT-5.6 Sol and Claude Fable, a case that "AI writing style collapse" is real, Brown University catching 50 students cheating with ChatGPT, and a Treasury comparison of AI-bubble risk to the dotcom era. Why it matters: One of the most-read recurring AI digests, useful for catching what mattered this week without reading everything yourself. // thezvi.substack.com

▸ Why the Rise of Open-Source AI Isn't Hurting Anthropic — Yet

Vercel spend data shows Anthropic still captures over half of AI platform spend even as token volume shifts to DeepSeek and other open-weight models — Opus 4.8 costs roughly 23x more per token than DeepSeek V4 Flash — because frontier and open models mostly aren't competing for the same workloads yet. Why it matters: A sharper, data-backed frame than the usual "open source is catching up" narrative. // TechCrunch