Top developments

Microsoft published the pre-training report for MAI-Base-2, a 2.8T total / 80B active MoE trained on 50T tokens across 16,000 GB200s over 38 days. Closed weights: the report is public, the model goes only to trusted partners via Foundry, and it is currently in post-training. The report claims 4.2 to 12.3x better pre-training FLOPs-efficiency than DeepSeek-V4-Pro-Base from scaling-law fits, a number to treat skeptically until it reproduces, and the model pairs with the Maia 300 inference chip slated for production in 2027. The scale is the story: 50T tokens and 16k GB200s puts this in the frontier training-run category, and publishing the report while keeping the weights closed is Microsoft playing the transparency game on its own terms. Worth reading for the large-scale MoE recipe details.

Meta released Muse Spark Open, an 816B total / 42B active open MoE with a 1M context window and native multimodality (text, image, video and audio in; text out). The headline claim is that it "runs on two Macs", but the fine print (the NVFP4 build, and the 512GB M5 Ultra Mac Studio hardware that is not on sale yet) means the consumer-hardware framing deserves skepticism until independent verification. Benchmarks are good but not dominant, with Meta itself billing it as the most capable US open-weight model, and the license reportedly still limits how large hosted businesses can offer it. An 816B MoE at 42B active is a serious open model; the two-Macs claim is marketing until proven.

NVIDIA's Nemotron 3.5 Super VL is a 120B total / 12B active open vision-language model shipped with open training recipes, datasets, and RL environments in NeMo Gym. The pitch is throughput: a claimed 5x speedup on single-GPU agentic tasks versus open peers. The open recipes and RL envs are the more important part: NVIDIA is commoditizing the post-training stack, which lowers the barrier for everyone building on open VLMs. The throughput claim needs independent verification, but efficiency as a headline feature is the right direction.

Physical Intelligence opened the Pi API in beta, a fine-tuning service for its robot foundation models: upload demos, fine-tune on managed infrastructure, run inference through a hosted gateway (pi0 and pi0.5 remain Apache 2.0 on GitHub). This is the foundation-model-as-API pattern applied to robotics: instead of training your own VLA, you fine-tune theirs. The beta is waitlist-gated, and the showcase examples (Almond's Axol robot deburring spacers, Actor's autonomous wheel loader) are the company's own, so independent validation is pending. If the base model is good, this could meaningfully accelerate robotics development; if not, it is just another API. The pattern matters either way.

A critical vulnerability (CVE-2026-111322, CVSS 9.8) hit the Archon coding agent: missing authentication on its /api/* control API, which binds to 0.0.0.0 by default, lets anyone with network access drive the coding agent (which runs with bypassPermissions) and execute commands on the host as the Archon user. A critical flaw in a coding agent is exactly the attack surface everyone warned about: agents that write and execute code are high-value targets. There is no patch yet, so if you are running Archon, do not expose the control API to any network you do not fully trust. This will not be the last such CVE in agentic coding tools.

Frontier AI labs

MLExpert is a 35B MoE built specifically for ML engineering tasks, trained with WIRL (Windowed Interaction RL), reaching a 72.7% medal rate on MLE-Bench Lite. Domain-specific models for ML engineering are an interesting bet: the generalist frontier models are already strong at coding, so a specialist needs to be meaningfully better at the ML workflow to justify itself. Worth a look for the training methodology.

Mercor's accounting study has frontier models now faster and more accurate than junior accountants on medium-length, well-defined accounting tasks, with Opus 5 scoring 100% across 20 attempts against a roughly 37% human average on its APEX-Accounting tasks. The gap is enormous, which either means the eval is too narrow or accountants should be worried. Most likely the former: Mercor itself notes the tasks test what AI is best at (detail-oriented instruction following), not the whole job, which includes judgment-heavy, unstructured work like client communication. Still, 100% versus 37% is a striking number even with caveats.

AV / embodied

Volvo expanded its NVIDIA partnership around the Hyperion sensor and compute architecture, Halos OS, and the Alpamayo driving software stack, targeting gradual introduction of eyes-off and mind-off driving, with no dates or launch model named. The details are thin, but Volvo deepening its NVIDIA relationship continues the trend of OEMs consolidating around NVIDIA's platform. Worth watching for what it signals about Volvo's in-house autonomy ambitions.

Waymo signed a binding framework for fully autonomous ride-hailing in Toronto, targeting first paid rides in 2028 with Element Fleet Management handling fleet operations: a meaningful regulatory milestone for international expansion, though Ontario's current pilot regulation is revoked in October 2027, so the 2028 date depends on rules that do not exist yet. Separately, Waymo opened rides to everyone in San Antonio, San Diego, and Tampa on October 14. The three-city open access is the scaling story: Waymo is moving from gated pilots to general availability in multiple markets simultaneously.