More from Soumith Chintala | Flash News

Thinking Machines: Releases Inkling-Small Model

Thinking Machines drops Inkling-Small with 276B parameters and 12B active, matching Inkling performance at quarter size for major AI industry impact. (Source)

Share 07-30-2026 17:55

Kimi K3: 2.8T Parameter Model Ships With 1M Context

Kimi K3 launches at 2.8 trillion parameters, 1 million context and native multimodal support, delivering 6.3x faster decoding via Kimi Delta Attention. (Source)

Share 07-16-2026 23:41

Modal: DFlash Speculator Achieves 67% Throughput Gain

Modal DFlash speculator delivers 67% higher throughput than MTP for Inkling on SGLang, accelerating AI inference speed optimization. (Source)

Share 07-15-2026 21:06

Bridgewater: Fine-Tuned Model Beats Frontier LLMs

Bridgewater fine-tunes Tinker model for financial news, delivering higher accuracy at lower cost than frontier LLMs. (Source)

Share 06-30-2026 19:27

Qwen 3.5 Launches with Hybrid Linear Attention and Vision Input

According to @soumithchintala, Qwen 3.5, developed by Alibaba Qwen, has been released and is now available on Tinker. The update introduces hybrid linear attention, which allows for extended context windows, and includes native vision input capabilities. This advancement is expected to significantly enhance AI application performance, particularly in tasks requiring long-context processing and multimodal inputs. (Source)

Share 03-06-2026 22:29

Soumith Chintala Announces Angel Investment in RadixArk as SGLang Team Scales Open AI Infrastructure

According to Soumith Chintala, he has joined as an angel investor in RadixArk and highlighted the differentiated culture built around SGLang and the team’s ambition. Source: https://twitter.com/soumithchintala/status/1998926753079935320. Ying Sheng stated that RadixArk is being built by many core SGLang developers from the LMSYS ecosystem, with SGLang started in summer 2023 and made public in January 2024. Source: https://x.com/ying11231/status/1998079551369593222. He added that growing demand for SGLang pushed the effort to be supported by RadixArk, whose mission is to make frontier-level AI infrastructure open and accessible, spanning schedulers, compilers, serving engines, and training pipelines. Source: https://x.com/ying11231/status/1998079551369593222. The thread referenced DeepSeek inference optimizations within SGLang and significant community contributions. Source: https://x.com/ying11231/status/1998079551369593222. No token, blockchain integration, or broader fundraising details beyond the angel commitment were disclosed, and the official site is radixark.ai. Source: https://x.com/ying11231/status/1998079551369593222 and https://twitter.com/soumithchintala/status/1998926753079935320. (Source)

Share 12-11-2025 01:24

AxiomProver Solves 9 of 12 Putnam Problems via ThinkyMachines Tinker: AWS-for-AI Signal for Infrastructure Traders

According to @soumithchintala, Axiom began four months ago and achieved notable Putnam results by bootstrapping its infrastructure on ThinkyMachines Tinker, highlighting a scalable AI infrastructure approach (Source: @soumithchintala on X, Dec 11, 2025). Axiom reported that its AxiomProver autonomously solved 9 of 12 Putnam problems in Lean, improving from 8 of 12 at 3:58 pm PT to 9 of 12 by noon the next day, which would have ranked first among roughly 4,000 participants last year and within Putnam Fellow range in recent years (Source: @axiommathai on X, Dec 10–11, 2025). @soumithchintala characterized this as an early proof-point that Tinker could play for AI research labs a role similar to AWS for startups in the 2010s, underscoring an infrastructure scalability narrative relevant to execution benchmarks (Source: @soumithchintala on X, Dec 11, 2025). (Source)

Share 12-11-2025 00:01

AI Text-to-Video Leaderboard 2025: Soumith Chintala Flags Whisper Thunder Poised To Topple VideoGen, Traders Monitor Momentum

According to Soumith Chintala, VideoGen is about to be toppled again with Whisper Thunder hinted as the challenger, referencing the artificialanalysis.ai text-to-video leaderboard for context, source: Soumith Chintala on X; artificialanalysis.ai text-to-video leaderboard. The linked leaderboard is the cited venue tracking text-to-video model standings, signaling a potential shift at the top that traders can verify directly on the page before acting, source: artificialanalysis.ai text-to-video leaderboard. The sources do not provide any crypto market or token-specific impact or price data related to this development, source: Soumith Chintala on X; artificialanalysis.ai text-to-video leaderboard. (Source)

Share 11-26-2025 14:53

Google Gemini 3 vs GPT-4: Soumith Chintala Flags Overwhelming Progress; Trading Watch on AI Equities and Crypto AI Narrative

According to @soumithchintala, Gemini 3 represents a sudden and overwhelming leap that feels closest to the GPT-4 moment, with Google looking invulnerable given its stack of TPUs, Android, and Chrome, while Anthropic quietly dominates code, though the race is not over (Source: @soumithchintala on X, Nov 23, 2025). For trading, this commentary signals a perceived leadership tilt toward Google's AI ecosystem and code-strong models, a sentiment input that market participants can track across AI-exposed equities and the AI-crypto narrative (Source: @soumithchintala on X, Nov 23, 2025). (Source)

Share 11-23-2025 17:34

Sunday Robotics Design Highlights: 2 Standout AI Robotics Details for Traders — Pinky Opens Dishwasher, In‑Hand Reordering

According to @soumithchintala, Sunday Robotics is praised as well-designed with specific details including a pinky opening a dishwasher and in-hand reordering of objects, source: @soumithchintala on X, Nov 21, 2025. The post expresses excitement and congratulates @tonyzzhao and @chichengcc, signaling positive attention toward Sunday Robotics from the author, source: @soumithchintala on X, Nov 21, 2025. The source includes no information on release timelines, funding, partnerships, pricing, or any cryptocurrency or token plans, so for trading purposes it constitutes a sentiment-only signal without quantifiable catalysts, source: @soumithchintala on X, Nov 21, 2025. (Source)

Share 11-21-2025 13:42

Soumith Chintala Leaves Meta (META), Steps Down From PyTorch on Nov 17: 90%+ AI Adoption, Exascale Training, 2025 Roadmap

According to Soumith Chintala, he will step down from PyTorch and leave Meta on November 17, 2025 (source: Soumith Chintala on X, Nov 6, 2025). He stated that PyTorch now handles exascale training, powers foundation models, is taught widely, and is in production at virtually every major AI company with 90%+ adoption in AI (source: Soumith Chintala on X, Nov 6, 2025). He emphasized a prepared succession bench—Edward, Suo, Alban, Greg, John, Joe, Jana, Jason and others—saying the project no longer depends on him and is truly resilient (source: Soumith Chintala on X, Nov 6, 2025). He pointed to a coherent PyTorch Conference product story and said the 2025 product lineup and execution will evidence continued strength (source: Soumith Chintala on X, Nov 6, 2025). He noted the PyTorch Conference now draws around 3,000 attendees where market-moving deals are brokered, underscoring deep industry integration (source: Soumith Chintala on X, Nov 6, 2025). He added that he is pursuing a small new venture outside Meta while staying involved with the community, including filing issues (source: Soumith Chintala on X, Nov 6, 2025). (Source)

Share 11-06-2025 18:36

Meta (META) AI Shake-Up: PyTorch Leader Soumith Chintala Steps Down After 8 Years, Citing 90%+ Adoption and Exascale Training

According to @soumithchintala, he will step down from leading PyTorch and leave Meta on November 17 after nearly eight years guiding the framework, stating he took PyTorch from nothing to over 90% adoption in AI, which is directly relevant for traders tracking AI infrastructure exposure at Meta (META) and its ecosystem partners, source: Soumith Chintala on X, Nov 6, 2025. He noted PyTorch now handles exascale training, powers foundation models, is in production at virtually every major AI company, and is widely taught in academia, highlighting the framework’s central role in enterprise AI stacks that investors monitor for continuity and roadmap execution, source: Soumith Chintala on X, Nov 6, 2025. Chintala emphasized leadership continuity, naming Edward, Suo, Alban, Greg, John, Joe, and Jana as ready, with culture carriers Greg, Alban, Ed, Jason, and Joe at the decision table and Suo, John, and Jana joining, signaling an orderly transition for a critical layer of AI software, source: Soumith Chintala on X, Nov 6, 2025. He added that the 2025 PyTorch product lineup and execution should provide evidence of the project’s resilience, offering a tangible near-term milestone for market watchers to track, source: Soumith Chintala on X, Nov 6, 2025. He described his role as leading the software layer that powers the entire AI industry and said every major AI company and hardware vendor are on his speed dial, underscoring the ecosystem significance of this leadership change for traders assessing AI supply chains, source: Soumith Chintala on X, Nov 6, 2025. (Source)

Share 11-06-2025 18:28

Soumith Chintala: PyTorch on Apple Mac Studio Lags NVIDIA; Meta Engineers Carry MPS — Trading Takeaways for AAPL and NVDA in 2025

According to Soumith Chintala, Apple's actual engineering time on PyTorch support has not given him confidence that the PyTorch Mac experience will get close to NVIDIA's any time soon, if ever, source: Soumith Chintala on X, Oct 16, 2025. According to Soumith Chintala, Meta engineers are doing a large share of the heavy lifting to improve the MPS backend and feel responsible for the Mac experience, while Apple's priorities and engineering hours fluctuate, source: Soumith Chintala on X, Oct 16, 2025. According to Soumith Chintala, PyTorch has over 90% AI market share and Apple must prioritize full PyTorch software support if it wants Mac Studio to be an AI development box rather than mainly an inference machine, source: Soumith Chintala on X, Oct 16, 2025. According to Soumith Chintala, the NVIDIA stack remains the reference for PyTorch training quality versus Apple's current MPS pathway, which is a trading-relevant signal for relative AI development readiness between NVDA and AAPL ecosystems, source: Soumith Chintala on X, Oct 16, 2025. According to Soumith Chintala, he did not mention cryptocurrencies such as BTC or ETH, indicating no direct crypto market impact is stated in his post, source: Soumith Chintala on X, Oct 16, 2025. (Source)

Share 10-16-2025 15:41

NVIDIA DGX Spark: Compact CUDA Dev Machine Highlights Software Moat — What It Means for NVDA Stock and AI Crypto Sentiment

According to @soumithchintala, NVIDIA’s DGX Spark is a compact CUDA development machine designed to sit on a desk with sufficient memory to handle very large parameter counts, prioritizing ease of development rather than peak speed or top benchmark performance; source: @soumithchintala on X, Oct 15, 2025. According to @soumithchintala, DGX Spark is positioned for a smooth workflow handoff: develop locally, then transfer the final training run to H200/B200 systems, deploy robotics policies on Jetson, and ship inference across vendors including NVIDIA, Apple, and AMD, indicating portability across the ecosystem; source: @soumithchintala on X, Oct 15, 2025. According to @soumithchintala, the statement that “NVIDIA wins because it’s a software company” underscores a software-led moat and developer lock-in narrative that traders often monitor when assessing platform durability for NVDA; source: @soumithchintala on X, Oct 15, 2025. According to the source, NVIDIA highlighted DGX Spark on its official X channel, corroborating product positioning and availability signals that equity traders may track for NVDA and peers; source: NVIDIA on X, post ID 1978200877983814091. According to @soumithchintala, explicit mention of cross-vendor inference pathways (NVIDIA/Apple/AMD) flags potential developer reach beyond a single hardware stack, a detail that stock traders in NVDA, AMD, and Apple may watch and that crypto market participants in AI-related narratives may track for sentiment around GPU-driven workloads; source: @soumithchintala on X, Oct 15, 2025. (Source)

Share 10-15-2025 14:47

Nvidia H100 $340B Estimate for OpenAI’s 10GW Build Raises NVDA Pricing, Margin, and Capex Signals

According to Soumith Chintala, a 10GW AI build equates to roughly $340B of Nvidia H100s at $30,000 per GPU, assuming 20% of power is reserved for non-GPU components (source: Soumith Chintala on X, Sep 23, 2025). He further estimates a 30% volume discount would reduce OpenAI’s outlay to about $230B (source: Soumith Chintala on X, Sep 23, 2025). Chintala contrasts the discounted scenario with a hypothetical where OpenAI pays full price and Nvidia reinvests the implied $100B delta into OpenAI equity, highlighting how deal structure could shift realized revenue versus strategic upside for NVDA (source: Soumith Chintala on X, Sep 23, 2025). For trading, the two price paths frame NVDA sensitivity to pricing power, gross margin mix, and potential strategic financing; if the 10GW plan referenced by OpenAI Newsroom on X guided actual orders, the magnitude implies a multi-hundred-billion-dollar pipeline that would be material for semis and AI-infrastructure equities while reinforcing AI-compute scarcity narratives that can influence crypto-adjacent GPU plays (sources: Soumith Chintala on X, Sep 23, 2025; OpenAI Newsroom on X as linked by Chintala). (Source)

Share 09-23-2025 12:26

Soumith Chintala Reveals Shift from VFX to Vision and ML Research: Implications for AI-Powered Crypto Trading

According to Soumith Chintala, his transition from VFX artistry to vision and machine learning research was driven by the pursuit of building intelligent agents capable of creative tasks, as shared via his Twitter account. This shift highlights the evolving landscape of AI talent moving into machine learning, a trend that is accelerating advancements in AI-powered trading algorithms and analytics. For cryptocurrency traders, such advancements can lead to more sophisticated quantitative strategies and real-time market insights, enhancing decision-making and risk management in the crypto markets (source: Soumith Chintala Twitter). (Source)

Share 08-04-2025 11:12

PyTorch Out-of-the-Box Model Training Continues Despite Infrastructure Failures: Impact on Crypto AI Trading

According to @data_and_ai, out-of-the-box PyTorch models continue training even when the underlying infrastructure experiences failures, raising concerns about model reliability and consistency in AI-driven crypto trading systems (source: @data_and_ai). This persistent training behavior could result in unreliable trading signals for cryptocurrencies like BTC and ETH, potentially increasing risk for algorithmic traders relying on AI-powered strategies. (Source)

Share 06-20-2025 18:59

Vibe-Coding Setup Revolutionizes GPU Kernel Development: Implications for Crypto & AI Markets

According to Soumith Chintala, the new Vibe-coding setup for GPU programmers, highlighted by @anneouyang, offers a breakthrough authoring experience that could set a new standard for custom GPU kernel development (Source: Twitter/@soumithchintala, June 14, 2025). This innovation enables developers to accelerate machine learning and AI computation, which directly impacts crypto mining efficiency and on-chain AI protocol performance. Traders should monitor related GPU and AI hardware stocks, as well as crypto assets reliant on high-performance computation, as these advancements may drive increased demand and price volatility in the crypto market. (Source)

Share 06-14-2025 01:38

RunwayML AI Film Festival 2025: Growing Influence of AI in Creative Industries and Crypto Market Impact

According to @soumithchintala, the RunwayML AI Film Festival is returning for its third edition in 2025, underscoring the expanding adoption of AI-generated content in the film industry (Source: Twitter/@soumithchintala). For crypto traders, this event signals increased institutional and retail interest in AI-related crypto tokens, such as $RNDR and $FET, which have shown positive price action following major AI industry events (Source: CoinGecko). Traders should monitor trading volumes and sentiment shifts around AI and creative industry tokens, as these may see volatility and upward momentum in response to heightened AI sector visibility. (Source)

Share 06-06-2025 00:18

PyTorch Foundation Expands as vLLM and DeepSpeedAI Join: Impact on AI and Crypto Markets in 2025

According to @soumithchintala on Twitter, the PyTorch Foundation is becoming an umbrella organization for leading AI open-source projects, with @vllm_project and @DeepSpeedAI joining as the first two members (source: Twitter, May 7, 2025). This consolidation of high-quality AI projects under PyTorch is expected to accelerate innovation and interoperability in the AI sector. For crypto traders, enhanced AI infrastructure can drive demand for blockchain-based AI solutions and related tokens, potentially increasing trading volume in AI-linked crypto assets as developers leverage open-source tools for decentralized applications. (Source)

Share 05-07-2025 14:24