Cognition released SWE-2, a coding model built on Moonshot AI's 2.8-trillion-parameter Kimi K3, which scores 50.0% on the company's FrontierCode 1.1 Main benchmark — within one point of Anthropic's Fa…
Cognition launched SWE-2, a coding model that scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while costing 64% less, and is available starting today in Devin Desktop and CLI. SWE…
DeepSeek launched V4.1-Flash on September 10, a 552-billion-parameter Mixture-of-Experts model with a 1-million-token context window that activates roughly 8 billion parameters for input and 16 billio…
DeepSeek released its V4.1 Flash model on Thursday, claiming it outperforms its previous flagship V4 Pro on coding, cybersecurity and autonomous agent benchmarks while cutting inference costs. The mod…
A study posted to arXiv (2609.09363v1) found only a weak context-memory conflict in large language models, with counterfactual RDF triple inputs scoring just 0.05 lower on a 1-5 faithfulness scale tha…
Moonshot AI has released Kimi K3, a 2.8 trillion parameter foundation model featuring a proprietary KDA hybrid linear attention mechanism and a 1 million token context window. The model demonstrates l…
A new analysis from Vals AI, an independent AI benchmarking company, found that longer, more complex AI tasks can have an environmental impact 10,000 times greater than simple queries, with building a…
A developer's v1.1 experiment reran reasoning-prefill tests using GPT-5.5 Pro as the teacher across 45 problems, finding that Qwen3.8 A95B's unigram source recall jumped from 33.92% to 54.50% (+20.58 …
A new web tool lets users estimate the energy and CO2 emissions of their AI usage, using modeled coefficients from public AgentX benchmarks on NVIDIA B300 hardware. The calculator, posted on Hacker Ne…
GitHub announced on September 3 that four GitHub Copilot models — Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 — will be retired on October 2, giving developers 23 days to m…
Together AI published a guide to the open source AI stack, detailing the components developers need to move from closed to open models, including models, inference, gateways, harnesses, and tools. The…
Argonaut Labs AI's Deltafin fork runs the full, never-pruned 2.8-trillion-parameter Kimi K3 model on a MacBook Pro at 0.2901 tokens per second (3.447 seconds per token), streaming from four SSDs, achi…
Magic reported that its pretraining recipe is now more than 10x more compute-efficient than leading open-weight base models, matching DeepSeek V4 Pro Base with roughly 50x fewer FLOPs — about half of …
Goldman Sachs is adopting a multi-model AI strategy that incorporates open-weight and open-source models alongside proprietary systems, pushing back against a default reliance on closed providers like…
A developer's practical review highlights four AI agent harnesses—Pi, OMP, OpenAI's Codex, and Zcode—used daily for coding and automation, with Pi praised as a minimalist open-source option supporting…
WorkBuddy, a desktop AI agent app, is offering free access to Tencent's Hi-4 Preview model for two weeks from launch, allowing users to try the 770 billion parameter mixture-of-experts model without p…
Tencent released Hunyuan Hi-4 Preview, a 770B-parameter mixture-of-experts model with 49B active parameters and a 1M-token context window, under the Apache 2.0 license. The model scores 92.3 on GPQA D…
Marvin, an AI agent from FutureHouse, scored 99/100 on Genentech's CompBioBench computational biology benchmark, with Marvin Lite scoring 95/100, outperforming GPT-5.6 Sol xhigh entries. The results h…
Meta's Muse Spark 1.3 scored 72.93% on the HTB-Challenger Benchmark, up from 50.41% for Muse Spark 1.2, according to a blog evaluation of the model on Hack The Box security challenges. The new version…
Moonshot AI's Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with 104 billion activated parameters per token and a 1,048,576-token context window, has brought attention to the debate betwe…