Jensen Huang resigns after 33 years at NVIDIA; Gemini 4 Hydrogen ships, GPU fingerprinting claims Chinese models trained on Blackwell
Latent Space · rss · 2026-11-13
The headline of this Latent Space AINews: Jensen Huang has stepped down as NVIDIA CEO, president, and board member after 33 years. It follows an NYT report of an affair with an accused Chinese spy and board preparations to replace him; a second affair was reported before the resignation. CFO Colette Kress (since 2013) becomes interim CEO ahead of what will likely be the most-watched earnings call in NVDA history. There's also debate over evidence that Chinese models were trained on Blackwell GPUs they shouldn't have access to.
Models
- Gemini 4 Hydrogen: Google's speed-first tier, claimed 2.5x faster than 3.8 Flash; $1.50/$7.50 per M tokens, 95% cached-input discount, 1M context, four thinking levels. Official: 57% on Terminal-Bench 4.0, 5x fewer fabricated claims. Artificial Analysis: AA index 54 at xhigh, 8% hallucination rate (vs 43% Sonnet, 52% Sol), but Terminal-Bench 4.0 only 54% vs Sonnet's 64%. Users report xhigh burning 131K thinking tokens / $1.08 without answering; Simon Willison notes it's effectively only a 25% discount vs Carbon/Argon
- Inkling 2 Small: Thinking Machines' Apache 2.0 release, 594B-total/17B-active MoE with image input and 1M context; novel components include attention residuals, gated attention, dynamic attention sink; up to 630 tok/s per user on one GB200 node via vLLM; #1 US open model on Code Arena WebDev (1647)
- Edge0-145B/120B: vendor-claimed on-device models hitting 86.5% on Terminal-Bench 2.1 (vs 78.2% for Opus 4.6) in 29GB at 75+ tok/s
- Perplexity Decider v1.3: open-weight decision/extraction models adding span-offset "locate" for NER/PIE/QA; claimed 6.54 points more accurate than production classifiers
Checkpoint forensics & Chinese compute
Will Depue claims weight initialization patterns fingerprint the training GPU: six Chinese models including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3 were allegedly initialized on Blackwell (four on Grace Blackwell); a B300 reproduced DeepSeek-V4-Flash's starting weights bit-for-bit; largest runs estimated at just 2-3K Blackwells. Critics note that's 6MW vs China's 24GW installed compute, suggesting export controls are working; other analyses place DeepSeek V4.1 Flash and LongCat on Ascend chips.
More from Models
- After two months daily-driving Claude Code, one reviewer says Opus 5.5 beats GPT everywhere except computer use — kimmonismus · 2026-11-13
- HF open-sources Ettin rerankers: six sizes, SOTA, full training recipe — antoine_chaffin · 2026-11-13
- LightOn open-sources DenseOn: SOTA single and multi-vector retrieval models — antoine_chaffin · 2026-11-13
- Voice Code Bench adds 8 ASR models as ElevenLabs Scribe V3 takes the top spot — lukeharries · 2026-11-13
- Claude Hit by Elevated Errors Across claude.ai, Claude Code and API Platform — ClaudeAI-mod-bot · 2026-11-13
- Claude is down again; user has Astra and Sol pick up the threads — paul_cal · 2026-11-13