Model launchesPrismML shrinks LLMs to run locally — Bonsai 2 27B compresses Qwen3.8 to 5.9 GBPrismML, a Caltech-founded startup led by Babak Hassibi, released Bonsai 2 27B, a compressed version of Alibaba’s Qwen3.8 27B that fits into 5.9 GB of storage — a roughly 9–10× reduction in memory with about 98% of the original’s benchmark performance.3 min read
Model launchesOpenAI launches 'Astra for Law'—GPT‑6 Astra tailored for law firms and legal techOpenAI introduced Astra for Law, a specialized legal foundation built on GPT‑6 Astra that combines a legal search index, custom instructions, and integrations for law firms and legal‑tech companies.5 min read
Model launchesGoogle launches Gemini 3.8 Live speech‑to‑speech models and a browser test UIGoogle has released two new speech-to-speech models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, and developers have already built a browser-based test interface for them.2 min read
Model launchesNeon: Menlo Park high-throughput materials science lab and open model to surpass GPT-6 AstraA high-throughput materials science lab and model-loop system set up in Menlo Park produces fresh experimental data, which were used to further train an open-source model via mid-training and reinforcement learning; using 1,300 H200 GPUs, the model named Neon exceeded GPT-6 Astra on an analysis benchmark.1 min read
Model launchesDeepSeek’s lean model supports Sutskever’s claim that scaling returns changedIlya Sutskever left OpenAI to found SSI and has argued since November that the industry’s focus on ever-larger models and scaling is reaching diminishing returns.2 min read
Model launchesSalesforce and Nvidia unveil Koa: an enterprise-focused, open-weight reasoning modelSalesforce and Nvidia introduced Koa, a reasoning-focused AI model post-trained from Nvidia’s open-weight Nemotron to handle sales, marketing, and customer-support tasks without ingesting customer data.3 min read
Model launchesEpoch Brief roundup: Huawei’s chip roadmap, Nvidia’s GDP omission, GPT-6 Astra milestonesThe Epoch Brief covers new research on Huawei’s Ascend roadmap and why the company is unlikely to match Nvidia this decade, an estimate that US GDP understates Nvidia’s contribution by about 0.3 percentage points, and architectural differences between OpenAI’s GPT and Anthropic’s Claude at long context lengths.5 min read
Model launchesNASA and IBM release Lunar Foundation Model to simplify lunar data analysisOn September 10, NASA and IBM published the Lunar Foundation Model on Hugging Face under an Apache-2.0 license.2 min read
Model launchesOpenAI pauses Pro subscriptions as demand for GPT-6 Astra overloads systemsOpenAI temporarily halted sales of its highest-tier ChatGPT Pro subscription after intense user demand for the newly released GPT-6 Astra strained its infrastructure.2 min read
Model launchesAnthropic's Claude models enter Microsoft Foundry and Microsoft 365 Copilot (public preview)On November 18, 2025, Microsoft and Anthropic expanded their partnership to make Claude Sonnet 4.5, Haiku 4.5 and Opus 4.1 available in public preview through Microsoft Foundry and to integrate Claude into Microsoft 365 Copilot features such as the Researcher agent and Agent Mode in Excel.3 min read
Model launchesDeveloper reverted to the GPT-5.6 and Fable 5.1 combo, criticism of GPT-6-Astra's mood swingsA developer reverted to using the GPT-5.6 and Fable 5.1 pairing after experiencing problems with the GPT-6-Astra model; they described GPT-6-Astra as simultaneously “the smartest and the dumbest,” criticizing its unpredictable behavior and technically correct but unreasonable solutions.1 min read
Model launchesOpenAI releases GPT‑Live‑1 in API for real‑time, interruptible voice interactionsOpenAI has launched GPT‑Live‑1 in its API, bringing full‑duplex, real‑time conversational voice capabilities to developers.3 min read
