Foundation Models & Reasoning
Core model architectures, training methods, chain-of-thought reasoning, and test-time compute scaling. The backbone of modern AI capabilities.
Key Benchmarks
Recent Papers
External Observers May See More Clearly: Cross-Model Span-Level Hallucination Detection in Large Language Models via Hidden State Probing
Kingshuk Gupta, Davide Buscaldi
Mingbird: A Local-First Agent Harness Enabling Small Open Models to Complete Real Tasks
Hao Wang, Ting Huang
TopK-Guided: Adaptive, Budget-Aware Activation Sparsity for Efficient LLM Inference
Mukund Agarwalla, Chih-Jen Lin
Sharpening Tax in Post-Training
Changdae Oh, Qi Zeng, Qi Qi +7 more
Hierarchical Continuous Diffusion Language Models
Hui Ren, Zihan Li, Chang Liu +2 more
GraphForge: Training Working Agents with Graph-Anchored Workspace Synthesis
Qisheng Su, Hanchen Wang, Guanru Zhu +2 more
Adaptive Reward Routing: Dynamic Multi-Reward Optimization for Joint Audio-Video Diffusion via Forward-Process RL
Songlin Yang, Xiaotong Zhao, Jiacheng Zhang +5 more
On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics
Julianna Piskorz, Antonin Berthon, Mihaela van der Schaar
ExpBoN: Exponential-Noise Best-of-n for Efficient Test-Time LLM Alignment
Yanxiao Liu, Sicheng Wan, Deniz Gündüz
Detecting Pretraining Data in Large Language Models from a Free-Energy Perspective
Chenye Ke, Zirui Liu, Qi Liu +4 more
Recent Milestones
Reflection Beam: open 501B MoE with 10K+ GB300s
On October 8, 2026 Capital & Compute published a deep analysis of Reflection AI’s Beam, a 501 billion parameter sparse MoE model with 23 billion active parameters whose open weights are due under Apache 2.0 later this month. Using Reflection’s own benchmark table, the piece finds Beam roughly matches Chinese open model GLM 5.2 on many tasks while trailing newer Chinese releases, and highlights that the model was trained with more than 100 million RL rollouts on over 10,000 NVIDIA GB300 GPUs.
Claude Haiku 5.5 outguns GPT-6 Luna on price tests
Anthropic released Claude Haiku 5.5 on October 8, 2026, pitching it as its cheapest and fastest small model for high volume tasks. The model undercuts Haiku 4.5 on price while claiming better performance than OpenAI’s GPT-6 Luna on several benchmarks, and is now available via Anthropic’s platform and major clouds. The launch completes the Claude 5.5 family just weeks before Anthropic’s anticipated IPO.
Falcon‑Emirati brings serious dialect AI
Researchers at Abu Dhabi’s Technology Innovation Institute have released Falcon-Emirati, a 7 billion parameter language model fine tuned to understand and generate Emirati Arabic, with details published on Hugging Face on October 6, 2026. The model captures local vocabulary, idioms and cultural references that standard Arabic and English centric LLMs often miss.
Moonshot AI hits $50B, targets 2027 HK IPO
Moonshot AI has closed its final private funding round at a valuation of about $50 billion and is preparing an IPO in Hong Kong in the first quarter of 2027, according to reports published October 6, 2026. The Beijing lab behind the Kimi chatbot aims to raise up to $5 billion in the listing.
Anthropic eyes $2T IPO, $518B compute buildout
Anthropic has circulated a draft IPO prospectus seeking a valuation between $1.8 trillion and $2 trillion, according to a report published October 6, 2026. The filing details at least $518 billion in long term compute and equipment commitments and warns that Anthropic’s own AI models could attempt to resist shutdown or manipulate users.
Mistral Large 4: 1T parameter open MoE
French startup Mistral AI has opened access to Mistral Large 4, a 1 trillion parameter mixture-of-experts language model that activates 49 billion parameters per query and will have its weights released later this month. The model launched October 6, 2026 in public preview on Mistral’s cloud and already ranks in the top tier on cybersecurity and vision benchmarks.
DeepSeek nears $12B round at $75B valuation
Chinese AI lab DeepSeek is close to raising at least 80 billion yuan (about $12 billion) in a new funding round backed by Tencent and battery giant CATL, according to reports on October 6, 2026. The round could grow to around 100 billion yuan and would value DeepSeek at a minimum of 500 billion yuan ahead of an early 2027 IPO.
Moonshot AI hits $50B, DeepSeek nears mega round
On October 6, 2026, multiple outlets reported that Chinese AI startup Moonshot AI has completed its final private funding round at a valuation of about 50 billion dollars and is preparing a Hong Kong IPO in early 2027 that could raise up to 5 billion dollars. The same reports say rival DeepSeek is close to securing between 80 and 100 billion yuan in new funding led by Tencent and CATL.
Trillion dollar AI buildout outpaces real productivity
A Reuters analysis published October 6, 2026, highlights that projected global AI data center and infrastructure spending could top $30 trillion by 2050, with US AI investment alone reaching about $9 trillion from 2025 to 2032. The piece notes that firms like Anthropic plan hundreds of billions in outlays despite limited evidence so far of broad productivity gains that would justify current valuations.
DeepSeek lines up $12B+ war chest for AI push
Bloomberg reporting summarized on October 7 and 8, 2026 says Chinese frontier lab DeepSeek is close to raising at least 800 billion yuan (about 12 billion dollars) in a new round led by Tencent and CATL, with the total potentially nearing 1 trillion yuan. The funding would reportedly value DeepSeek at roughly 5,000 billion yuan, around 75 billion dollars, ahead of a planned 2027 IPO and large scale data center buildout in Inner Mongolia.
China’s 15th Plan makes AI hardware a core pillar
On October 6, 2026, China Economic Daily detailed a new “15th Five‑Year Plan” for the electronic information manufacturing sector that emphasizes AI as a core growth driver. The plan calls for strengthening domestic AI chips and sensors, upgrading phones and PCs into AI terminals, and expanding smart devices like wearables, robots and automotive electronics.
Beam: 501B open model rivals Chinese leaders
Reflection AI unveiled Beam, a 501B-parameter sparse Mixture-of-Experts open-weight model, on October 5, 2026. The lab says Beam matches or approaches top Chinese open models on coding and reasoning tasks while using three to four times less inference compute.([reflection.ai](https://reflection.ai/blog/introducing-beam))
India readies $25B deep-tech AI and chips fund
On October 4, 2026, Channeliam reported that India is preparing a potential $25 billion funding pool for deep‑tech startups in AI, semiconductors, advanced manufacturing, drones and space. The pool would combine about $11 billion already committed via the Research Development Infrastructure Fund with matching capital and additional contributions from private investors.
Aleph Alpha drops 78B Kolibri-1 as open weights
Aleph Alpha released Kolibri-1 on October 3, 2026, a 78 billion parameter mixture-of-experts language model optimized for German and English with explicit reasoning modes and tool calling. The company published a detailed technical blog and model card, and made the full weights available on Hugging Face under an Apache 2.0 license.
SoftBank taps $11B junk bonds to double down on OpenAI
SoftBank Group launched a high-yield bond sale of about $11.1 billion on September 21, 2026 to help fund a $10 billion follow-on payment for its OpenAI investment and refinance an earlier bridge loan. Term sheets show $10 billion in US dollar notes and €1 billion in euro notes across multiple maturities, with pricing expected on September 24 and settlement on September 29.
OpenAI Targets $1.2T Valuation In New Round
The Information reports that OpenAI is in early talks with investors about a new private funding round that could value the company at $1.2 trillion or more. Other outlets say this would add roughly $350 billion in paper value since a March round that priced OpenAI at about $852 billion. ([theinformation.com](https://www.theinformation.com/newsletters/dealmaker/openais-next-round?utm_source=openai))
DeepMind spinoff Emulate targets $700m war chest
On September 17, 2026, Sifted and other European outlets reported that Emulate, a London AI startup founded by former Google DeepMind researchers, is in talks to raise up to $700 million at a post‑money valuation of around $3.7 billion. The round is reportedly being led by Index Ventures and Lightspeed, though terms are not yet final.
OpenAI eyes $1.2T pre‑IPO valuation
Citing the Financial Times and Bloomberg, multiple outlets reported by September 16, 2026 that OpenAI is in early talks with investors about a new funding round targeting a $1.2 trillion valuation. A detailed TMTPost analysis says the company now plans to postpone its IPO to 2027 while seeking to add roughly $350 billion in value on top of its March 2026 round.
Fitch warns of systemic AI investment crash
Cinco Días reported that Fitch Ratings has modeled a scenario in which an AI investment bubble bursts, sending Wall Street down 35 percent in six months and tipping the U.S. into recession in 2027. The analysis draws on earlier work by the BIS that highlights almost €1 trillion in AI related capital spending by a handful of major firms and warns that a correction could exceed the dot-com bust. The article was published from Madrid at 05:30 CEST on September 16, 2026.
China backs 10k AI SMEs in 'rainforest' push
On September 15, 2026, Xinhua Daily commentary republished by Sina highlighted China’s new “AI SME Entrepreneurship Support Plan (2026–2028),” aiming to foster a “rainforest‑style” ecosystem for AI startups. The plan targets over 10,000 new tech and innovative AI SMEs and more than 2,000 ‘little giant’ specialized firms within three years.