"The AI Pulse — 2026-10-23: A Flagship Built Without American Chips"

Open-weight flagships stopped being a Silicon Valley genre some time ago, and today the genre's two biggest new entries came from Chinese labs working entirely outside the US export-control fence — while, elsewhere, a 14.8B model learned to read genomes and a fight broke out over whether neural networks are even legible.

A 293B flagship on 10,000 domestic cards

At its 1024 Developer Festival, iFlytek open-sourced Spark-X2.6, a 293B-parameter mixture-of-experts model (30B active) with a 1M-token context window, trained end-to-end on a 10,000-card cluster of domestic Chinese accelerators — no NVIDIA, no workarounds. The company claims coding and agentic breakthroughs on homegrown silicon, which matters less as a benchmark entry than as proof that an entire frontier-model pipeline, from chips to weights, now runs outside the reach of US export controls. And it wasn't even alone: the same week Tencent dropped Hunyuan Hy3, a 295B MoE (21B active, 256K context) under Apache 2.0 with free commercial use, the third frontier-class open Chinese MoE in seven days. The economics of closed-API pricing keep getting squeezed from a direction no Western vendor controls. iFlytek release via People's Daily · Hunyuan Hy3 repo

"Chat with your genome," for the price of a dinner

Hugging Face released Carbon-X, a 14.8B Apache-2.0 model that takes raw DNA as input and answers questions in plain English, averaging 66.7% across five DNA benchmarks at roughly one-1,800th the inference cost of a frontier model. Because it runs on a single 80GB GPU, a hospital or university lab can screen variants without shipping patient genomes to anyone's API — Hugging Face estimates classifying ClinVar's 2.4M uncertain variants would cost about $300, versus $32K–$4M at frontier-lab rates. The model card is refreshingly honest that wet-lab validation doesn't exist yet, but as a first practical step toward genomics that small labs can actually own, it's a genuine open-science play. Carbon-X release

Are neural networks already legible?

Goodfire's Nick Cammarata argued this week that the field has misread its own neurons: about 90% of Qwen3-32B's neuron poles seem to have one dominant job, meaning polysemanticity may be far rarer than the "neurons are a tangled mess" consensus assumed. If he's right, mechanistic interpretability shifts from an untangling project to a reading project, and safety auditing of large models gets dramatically cheaper. The claim is contested — Neel Nanda and others urge caution, and the companion paper is still pending — but the debate itself is the signal: what safety engineers can and can't verify inside frontier models is suddenly an open empirical question again, not a settled pessimism. The essay

The resignations DeepMind can't gag

BBC Eye reporting documents a quiet talent bleed at Google DeepMind over the Pentagon contract: research engineer Sofia Liguori quit this week despite a contract gag order to say so publicly, following researcher Andreas Kirsch's October exit, both citing the lab's lack of visibility or veto over classified military use of its models under the "any lawful government purpose" clause. For a company whose AGI roadmap depends on retaining exactly these researchers, the question is no longer whether frontier labs will have military customers, but whether the people who build the models get any say in how they're deployed. BBC investigation

---

Watch-item for tomorrow: whether anyone runs independent benchmarks on Spark-X2.6's coding-agentic claims — sovereign-compute milestones deserve proof, not press releases, and Hy3's Apache-2.0 weights are already sitting in the repo waiting for a skeptical harness.