Hugging Face Releases Ripple-65B, Claims US Open-Weight SOTA in Agentic Coding

Key Takeaways

1. Hugging Face released Ripple-65B Preview, a 65B dense model pretrained from scratch, claiming a new US open-weight SOTA in agentic coding over several larger models.
2. Hugging Face reported 18,000 new science models added in 2026 to date, versus 6,100 in the same period of 2025.
3. Google expanded its eighth-generation TPU line with TPU 8i Max and TPU 8tx, the first time it has launched four systems in one generation within a year.
4. NVIDIA added Dynamo deployment recipes for four Korean models, claiming up to 2x inference performance on B200 GPUs.
5. A French AI bill passed the Senate unanimously, shifting proof burdens to model providers; critics including Mistral's Arthur Mensch raised feasibility and trade-secret concerns.
6. CXMT was described as the world's fourth-largest DRAM maker at roughly 10% global share, with two-thirds of business from phone memory.
7. MaxxBench results showed a model scoring about 74% on DeepSWE varying by 23 points on lightly altered tasks, prompting cautions about public benchmarks.

1. AI Models and Infrastructure

1. Hugging Face released Ripple-65B Preview, described as a 65B dense model pretrained from scratch that sets a new US open-weight SOTA on agentic coding, claiming it significantly exceeds the larger Muse Spark Open-816B, Beam-501B, Inkling-975B and Nemotron 3 Ultra-550B. The company also reported harness adapters trained over 1,900+ RL steps for LFM2.5-24B, LFM3 and Twil-8B, producing nine adapters and GGUF files; on a held-out data-analysis test set, LFM3 tool efficiency rose up to 43% and accuracy 15%, while Twil-8B tool-call efficiency rose up to 37%, with all reward-hacking attempts blocked by safeagents and open-sourced. Separately, AI for Science publications on Hugging Face grew to 18,000 new science models in 2026 to date from 6,100 a year earlier, a gap the publisher said it had initially underestimated. — via 1 2 3
2. Google expanded its eighth-generation TPU lineup with TPU 8i Max, aimed at ultra-long context and high rack compute density, and TPU 8tx, aimed at secure confidential inference; it said this is the first time it has introduced four different TPU systems in one generation within a year. Developers can run native PyTorch and Hugging Face models through TorchTPU without code rewrites, while open-source TPU Sync cuts KV transfer latency by more than 90% and Agent Compute Engine targets idle compute waste. — via 1
3. NVIDIA added Dynamo deployment recipes for SK Telecom's A.X K2, LG AI Research's K-EXAONE 2.0, UpstageAI's Solar Open 2 and Motif Tech's Motif-3, claiming up to 2x inference performance on B200 GPUs. At AI Day Seoul, attended by more than 2,000 developers, startups and researchers, Hyundai Motor Group was said to be exploring NemoClaw and OpenShell, AIMSafety integrated OpenShell for safer enterprise agent use, and LG AI Research opened API access to K-EXAONE-2.5 trained on NVIDIA Nemotron datasets. — via 1 2 3

2. Policy, Chips and Evaluation

1. A French AI bill, the Darcos law, passed the Senate unanimously on April 8: in civil litigation, once there are indications a work may have been used, the model provider is presumed to have used it and must prove otherwise. Opponents cited evidentiary standards, cost and trade-secret disclosure in proving that billions of tokens of training data exclude a given work, with Mistral's Arthur Mensch saying this is impossible without revealing methods. The same forwarded discussion said Mistral has released the 1-trillion-parameter open-weight model Large 4, trained on its own machines at Bruyères-le-Châtel, and is developing 1 GW of compute mainly in Europe and France; the forwarder worried the law would more easily hold accountable vendors training in France and proposed a uniform contribution on European revenue instead, and noted Spotify and Believe have signed paid AI-mixing licensing deals. — via 1
2. CXMT was described as about a decade old and the world's fourth-largest DRAM maker after Samsung, SK Hynix and Micron, with roughly 10% global share, a $430 billion valuation and a brief spell as China's most valuable company. The forwarded account said revenue grew more than sixfold year over year to $40 billion in the first nine months with a 50% gross margin, about two-thirds from phone memory and a small AI-compute share; founder Zhu Yiming studied at Tsinghua and Stony Brook, worked in Silicon Valley semiconductors and returned in 2005 before founding CXMT in Hefei in 2016, where the city government holds about 37% and absorbed a decade of losses. — via 1
3. MaxxBench, with 89 DeepSWE tasks altered only slightly in instructions or starter code, showed a model scoring about 74% on DeepSWE varying by as much as 23 percentage points, raising the question of whether models are good at coding or at DeepSWE. Hamel Husain argued against reading it as another benchmark for picking models by open/closed status or family, saying judgment requires private evals on one's own tasks and that all public benchmarks deserve healthy skepticism; a forwarded remark compared benchmarks to perfume—fine to smell, not to drink. — via 1 2 3

3. Research Directions and Product Signals

1. Forwarded research discussion distinguished JEPA from JEA by joint-embedding versus generative architecture, with the predictor needed only for asymmetric prediction such as corrupted views or time series, while both can add an action variable to become world models; the lineage listed Siamese nets, Barlow Twins, VICReg, SIGReg and DINO. A separate forward introduced LeHami, which brings Hamiltonian mechanics into JEPA using first-order equations for coordinates q and momentum p, kinetic-like K(p) and potential-like V(q,u) terms, and a symplectic leapfrog integrator to preserve phase-space volume and prevent latent collapse in long-horizon prediction. — via 1 2
2. Greg Brockman said frontier models paired with large amounts of test-time compute are becoming an engine of scientific discovery, while a forwarded post said a Google DeepMind project is witnessing a genuine phase transition in the speed of scientific discovery with demonstrations expected in coming weeks. Ethan Mollick drew an optimistic analogy from The Rise and Fall of American Growth, which dates a one-time leap in US living standards to 1870–1920 and judged in 2016 that such a "special century" was unlikely to recur, asking whether AI could deliver another and how to keep it on track. — via 1 2 3
3. Perplexity said Teammates usage has risen daily since launch and teased three releases this week: reference Teammates open to all, Teammates for financial research, and a simpler automation setup flow. Runway launched Runway Nights, a creative-community event series starting November 18 in New York with limited, application-based seats. Mistral AI said it used Voxtral Transcribe 3 to transcribe all main-stage talks at AI Engineer Paris in real time, feeding speaker names and agenda technical terms through context biasing to improve accuracy. — via 1 2 3 4