📡 The Week Ant Group Solves Spatial Perception at 1B Params, Tencent Ships a 295B Open Model

Ant Group's Robbyant open-sourced a 1B vision model that beats models 7x its size at dense spatial perception. Tencent dropped a 295B MoE under Apache 2.0. NVIDIA unified audio and text in one decoder without losing IQ. And Liquid AI cut "doom loop" rates in reasoning models by 90% with one training trick.

The Week Fable 5 Comes Back, China Closes the Gap, and Proof Assistants Get an Agent

Anthropic's Fable 5 is back online after 18 days dark. Zhipu's GLM 5.2 landed within 1 point of Opus 4.8 at a fifth of the cost. Mistral solved 587 of 672 Putnam problems with an open model. And three separate teams shipped agents that can read — your browser, your webpage, and your Fitbit. Here's everything.

The Week Governments Started Gating Frontier AI and Open-Source Filled the Vacuum

GPT-5.6 landed — then got restricted to 20 trusted partners at the White House's request. Meanwhile: DeepSeek made its own models 85% faster without a new GPU. Liquid AI shipped a 230M model that runs on a mobile NPU. Meta open-sourced a design system agents can actually read. Here's everything.

The Week Apple Released Container & Baidu Released Unlimited OCR

Apple dropped a containerization tool that runs Linux as lightweight VMs on Apple Silicon — no Docker, pure Swift. DeepReinforce open-sourced a coding model that matches closed frontier systems.

📡 AI INTELLIGENCE BRIEF: The Week PDFs Got a Schema, Kernels Got Faster, and Grok Went Autonomous

Datalab ships a 9B PDF-to-JSON extractor. Prime Intellect trains trillion-parameter MoE with RL. Sakana routes tasks across frontier LLMs through one API. xAI drops /goal for long-running autonomous coding. GPT-5.6 previewed, Noam Shazeer joins OpenAI. Here's everything from the last 4 days.

📡 AI INTELLIGENCE BRIEF: The Week AI Memory Got a Brain and Agents Got a Directory

28.4× Reduction in per-token attention compute at 1M context with MiniMax Sparse Attention — with zero quality loss vs. dense GQA. Prefill 14.2× faster. Decode 7.6× faster. On H800. Long-context just got dramatically cheaper to serve.

📡 AI INTELLIGENCE BRIEF: The Week the US Government Pulled the Best AI Model Ever Built.

The highest-scoring coding model ever shipped got disabled by government order. A 1M-token context window shipped with no benchmarks. The orchestration layer above coding agents just became open source.

📡 AI INTELLIGENCE BRIEF: The Week 3 Billion Parameters Beat 120 Billion. Efficiency Won.

95.0% Claude Fable 5's score on SWE-bench Verified — the highest ever recorded by any model. The previous record was 88.6%, held by Claude Opus 4.8 — released just two weeks ago. GPT-5.5 sits at 58.6% on SWE-bench Pro. Anthropic just lapped the field.

Inside: Xiaomi pushes 1T model past 1,000 tok/s → NVIDIA ships 550B agent MoE → Microsoft transcribes an hour of audio in 15 seconds → Google's RAG loop raises factuality 34%