More from Soumith Chintala | AI News

Laguna S 2.1 Debuts: 118B MoE Breakthrough

According to soumithchintala, Poolside’s Laguna S 2.1 packs 118B MoE with 1M context and runs on a single NVIDIA DGX Spark, with open weights on Hugging Face. (Source)

Share 07-22-2026 00:34

Kimi K3 Debuts with 2.8T Params, 1M Context

According to soumithchintala, Kimi K3 ships 2.8T params, 1M context, multimodal, faster decoding, and will open weights by July 27, 2026, per Kimi.ai. (Source)

Share 07-16-2026 23:41

Modal DFlash speculator boosts inference 67%

According to soumithchintala, Modal’s DFlash speculator for Inkling delivers 67% higher throughput and interactivity than MTP on SGLang endpoints. (Source)

Share 07-15-2026 21:06

Inkling Launches 975B Open Weights Multimodal Model

According to soumithchintala, Inkling debuts with 975B params, open weights, and native text image audio support on Tinker and Hugging Face. (Source)

Share 07-15-2026 18:15

Thinking Machines Unveils Decentralized AI Vision

According to @soumithchintala, Thinking Machines pushes personalization, human-in-the-loop, and decentralization to cut reliance on centralized AGI. (Source)

Share 07-10-2026 17:18

ThinkyMachines Unveils Personalization Playbook

According to soumithchintala, ThinkyMachines targets personalization, human-in-the-loop, and decentralization to reduce AGI platform dependence. (Source)

Share 07-10-2026 17:13

Bridgewater Fine-Tunes Model Beats Frontier Costs

According to soumithchintala, Bridgewater fine-tuned a model for financial news triage that outperforms frontier LLMs on cost and reliability. (Source)

Share 06-30-2026 20:18

Bridgewater Fine-Tunes Model Beats Frontier LLMs

According to soumithchintala, Bridgewater’s fine-tuned model ranks financial news better and cheaper than frontier LLMs, per Tinker and Thinking Machines. (Source)

Share 06-30-2026 19:27

Flourish AI Labs targets human-level efficiency

According to soumithchintala, Flourish AI Labs aims to match human sample efficiency and energy use, a shift that could reshape AI hardware and training. (Source)

Share 06-05-2026 13:02

Interaction Models Enable Live System Design Demos

According to soumithchintala, demos show Interaction Models co-designing systems, reading papers, and live fact-checking with a generative UI. (Source)

Share 05-13-2026 18:39

Thinking Machines Hires Supercomputing Engineers

According to @soumithchintala, Thinking Machines is hiring supercomputing engineers for real time models, Tinker, and large scale training in NYC and SF. (Source)

Share 05-12-2026 17:12

Thinky Interaction Models boost human AI bandwidth

According to @soumithchintala, Thinky previews real‑time Interaction Models that expand human AI bandwidth and collaboration, per Thinking Machines. (Source)

Share 05-11-2026 20:48

Test of Time LLM Debuts With Retro Benchmark Fun

According to @soumithchintala, Test of Time LLM offers a playful, retro-style benchmark link, highlighting community interest in evaluators. (Source)

Share 04-28-2026 13:25

Claude Boosts Enterprise Support Scale Analysis

According to @soumithchintala, Anthropic may scale account support via Claude or humans, while firms adopt multi AI with open harnesses for flexibility. (Source)

Share 04-27-2026 13:29

Jensen Huang Podcast Analysis: Ecosystem Strategy, Test-Time Compute, and Policy Levers in AI 2026

According to Soumith Chintala on X, Jensen Huang’s conversation with Dwarkesh Patel highlights that AI progress is driven by ecosystem dynamics, supply chain control, and incremental compute plus post-training advances rather than a single phase-change model event, as reported by Soumith Chintala. According to the podcast outline by Dwarkesh Patel, the discussion covered Nvidia’s supply chain moat, TPUs’ competitive threat, and export policy to China, underscoring business implications for chip vendors and hyperscalers. According to Soumith Chintala, a realistic baseline is that a state-of-the-art Chinese open-source model could gain three orders of magnitude more test-time compute with unpublished post-training techniques, implying competitive parity risks for Western firms and the need for layered policy interventions. As reported by Soumith Chintala, overzealous early regulation could harm U.S. competitiveness; instead, measured, continuous controls across the ecosystem—from chips and interconnects to software stacks—are recommended, creating opportunities in compliance tooling, inference optimization, and supply chain orchestration. (Source)

Share 04-20-2026 16:32

Thinking Machines acquires Workshop Labs: Analysis of human-in-the-loop AI strategy and 2026 growth outlook

According to Soumith Chintala on X, Workshop Labs is joining Thinking Machines to pursue AI that amplifies humans rather than replaces them, with links to the official announcement and blog post for confirmation. As reported by Workshop Labs’ blog, the team frames the deal around building human-in-the-loop systems that elevate expert workflows, indicating a product focus on copilots and decision-support tools for enterprises. According to the Thinking Machines announcement posted via X, new hires including Luke Drago and LRudL will contribute to applied AI capabilities, suggesting near-term offerings in data-driven copilots, retrieval-augmented generation, and workflow orchestration for knowledge-heavy teams. For businesses, the move signals opportunities to deploy AI copilots that integrate with existing data stacks, reduce time-to-insight, and maintain human oversight—key for regulated sectors such as finance, healthcare, and public services, as stated in the Workshop Labs blog post and associated X thread. (Source)

Share 04-13-2026 18:05

NVIDIA Backs Thinking Machines: 1GW Compute Partnership for Frontier Model Training – Latest Analysis

According to soumithchintala on X, Thinking Machines has partnered with NVIDIA to bring up 1GW or more of compute starting with the Vera Rubin cluster, co-design systems and architectures for frontier model training, and deliver customizable AI platforms; NVIDIA has also made a significant investment in Thinking Machines (as reported by the official Thinking Machines announcement at thinkingmachines.ai/news/nvidia-partnership/). According to Thinking Machines, the collaboration targets large-scale training efficiency and verticalized AI deployment, indicating near-term opportunities in AI infrastructure provisioning, GPU-accelerated training services, and enterprise model customization. (Source)

Share 03-10-2026 13:51

Qwen 3.5 Launch on Tinker: Hybrid Linear Attention, Long Context, and Native Vision Input – Latest Analysis

According to Soumith Chintala on X, four Qwen 3.5 models from Alibaba Qwen are now live on Tinker, introducing hybrid linear attention for extended context windows and native vision input support (source: Soumith Chintala; original post by Tinker and Alibaba Qwen). According to Tinker, this enables developers to deploy Qwen 3.5 variants for long-document reasoning and multimodal workflows with reduced memory overhead, improving inference efficiency and context handling for enterprise RAG, meeting transcription, and analytics use cases. As reported by Alibaba Qwen’s announcement referenced in the post, native vision input allows image understanding without extra wrappers, opening opportunities for e commerce visual search, industrial inspection, and content moderation pipelines. According to the cited posts, immediate availability on Tinker lowers integration friction for startups and enterprises seeking scalable long context LLMs with vision capabilities, supporting faster prototyping and cost efficient production deployment. (Source)

Share 03-06-2026 22:29

Anthropic CEO Issues Statement on Talks with US Department of Defense: Policy Safeguards and Model Access – Analysis

According to Soumith Chintala on X, Anthropic shared a statement from CEO Dario Amodei about discussions with the US Department of Defense, outlining how the company evaluates government engagements, sets usage restrictions, and preserves independent oversight; according to Anthropic’s newsroom post by Dario Amodei, the company will only provide model access under strict acceptable-use policies, red teaming, and alignment controls designed to prevent misuse, and it will not build custom offensive capabilities, emphasizing safety research, evaluations, and transparency commitments; as reported by Anthropic, the approach aims to balance national security cooperation with responsible AI deployment, signaling opportunities for enterprise-grade compliance solutions, safety evaluations as-a-service, and policy-aligned model offerings for regulated sectors. (Source)

Share 02-27-2026 12:56

Meta Open-Sources Llama 3.3: Latest Analysis on Model Access, Licensing, and 2026 AI Ecosystem Impact

According to @soumithchintala, the referenced announcement is “as wild as OpenAI dropping the open,” signaling a major shift in AI model access and governance. As reported by Meta AI’s model releases and industry tracking sources, Meta has continued to open-source advanced Llama versions under permissive licenses enabling commercial use, which contrasts with OpenAI’s closed distribution and suggests intensified platform competition for developers, inference providers, and edge deployment partners. According to Meta’s Llama license and release notes, open weights lower total cost of ownership for startups via on-prem and VPC inference, expand fine-tuning freedom, and accelerate vertical solutions in customer support, code assistants, multilingual RAG, and on-device AI. As reported by venture analyses and cloud benchmarks, this dynamic pressures cloud margins, drives optimized inference (AWQ, vLLM, TensorRT-LLM), and creates opportunities for model hubs, eval providers, and enterprise guardrail vendors. According to ecosystem data cited by model hubs and MLOps platforms, the business upside includes faster time-to-market for SMEs, sovereignty compliance in regulated regions, and new monetization for hosting, safety, and retrieval orchestration. (Source)

Share 02-25-2026 17:04