Morning Gradient: This Week in AI

Agentic coding at 1M tokens — Meta's Muse Spark 1.1 pairs a 1-million-token context window with parallel subagents built for orchestrating large, real-world codebases, and ships with a public, OpenAI-compatible Meta Model API preview that can script efficient paths instead of clicking through every step.

Meta blog

Claude, verified for K-12 classrooms — Anthropic's Claude for Teachers gives verified US K-12 educators free access to premium capabilities plus curriculum-aligned lesson help via Learning Commons, wrapped in FERPA-compliant data handling and a pilot with Detroit Public Schools.

Anthropic blog

xAI's newest 'smartest model yet' — Grok 4.5 arrives with claimed strengths in coding, agentic tasks, and general knowledge work — though the 'smartest model yet' title historically has a shelf life of about one press cycle. Benchmark figures and technical specifics are still thin.

xAI blog

Robot navigation with one camera — Mistral's 8B Robostral Navigate hits a 76.6% success rate on the R2R-CE benchmark using only a standard RGB camera — no depth sensors or LiDAR — and trains 22x more token-efficiently thanks to a prefix-caching attention trick.

Mistral blog

Gemini assistant for Indian classrooms — Built with India's Atal Innovation Mission, ATL Saathi uses Gemini 3.5 Flash to generate project ideas, infographics, and assembly instructions for Atal Tinkering Labs educators, rolling out in eight languages across 100 schools.

Google DeepMind blog

New scaling axis for LLMs — xHC expands Transformer residual streams into 16 parallel channels but only updates four at a time, letting 18B–28B models beat baseline downstream scores while matching the loss of standard training at two-thirds the compute — sixteen channels, four active, group-chat energy. A custom 'xHC-Flash' kernel keeps it practical at scale.

arXiv

2M-token RL on eight GPUs — LongStraw's architecture-aware execution stack pushes GRPO-based RL post-training past 2.1 million tokens on just eight H20 GPUs, narrowing the gap between the long contexts models can read and the much shorter ones they're usually trained on.

arXiv

Distilling the reasoning delta — On-Policy Delta Distillation trains student models on the specific reasoning gains a teacher picked up over its base version, rather than just its final outputs — a cleaner signal that beats conventional distillation across math, science, and coding benchmarks.

arXiv

Comment sections as attack vectors — Researchers show pretraining data can be poisoned through public comment sections far more easily than by editing curated sources like Wikipedia, and introduce HalfLife, a tool for estimating how much of that adversarial content survives web-crawl cleaning.

arXiv

Real-time rewards for diffusion models — A new continuous-time RL framework derives PPO and GRPO variants for discrete diffusion models, enabling intermediate rewards throughout the denoising process instead of only at the end — a step toward principled post-training for non-autoregressive models on math and coding tasks.

arXiv

Self-hosted LLM interface leader — Open WebUI remains the go-to TypeScript front end for local and API-based models, with RAG support, multi-model conversations, and a plugin system for extending it past the defaults. Currently at 54,000 stars.

GitHub

IBM's answer to messy PDFs — Docling converts PDFs, DOCX, and images into clean Markdown or JSON while preserving layout structure, sparing RAG pipelines from hallucinating because a table got mangled into a chaotic string of text. At 12,500 stars.

GitHub

A private assistant for your data — Khoj is a self-hosted personal assistant that searches your own notes, emails, and images in natural language, aiming to keep your 'digital brain' off someone else's server. Sitting at 21,000 stars.

GitHub

Controllable, less robotic speech — Parler-TTS gives developers fine-grained control over pitch and rhythm in text-to-speech, aiming for AI voices that sound less like a hostage negotiation and more like a person. At 6,200 stars.

GitHub

Still the agentic autonomy testbed — AutoGPT remains the reference project for autonomous task execution, breaking goals into sub-tasks for GPT-4 to chase down — the ambitious, if slightly chaotic, frontier of agentic AI, now at 168,000 stars.

GitHub