Bolt.new launched a mode called Forge on September 14, 2026 that gives every individual Pro subscriber up to 50 times more usage on four open-weight coding models — GLM 5.3, GLM 5.3 Flash, Kimi K3, an…
Mouse 0.1.0, an open source harness for long-running coding agents built on OpenCode, passed 25 of 30 tasks (83.3%) on FrontierHarness Eval with Kimi K3 on 2026-09-08, at a cost of $2.79 per pass and …
Bolt.new co-founder and CEO Eric Simons launched Bolt Forge on September 14th, giving paid individual users up to 50 times their usual AI-building allocation in exchange for opting into a training-dat…
Abacus.AI released three open-weight Smaug models — Smaug Flash on DeepSeek V4 Flash, Smaug Mini on Qwen3.8 27B, and Smaug Agentic on Kimi K3 — tuned for long-running agent loops and published on Hugg…
Moonshot AI released the weights for its Kimi K3 model on July 27, 2026, after announcing the model on July 16; independent measures place Kimi K3 within one generation of U.S. frontier models. The Tr…
Cognition released SWE-2, a coding model post-trained with reinforcement learning from Moonshot AI's 2.8T-parameter Kimi K3, scoring 50.0% on FrontierCode 1.1 Main — within 1 point of Fable 5.1 at 64%…
Cognition released SWE-2, a coding model scoring 50% on FrontierCode 1.1 Main — within one point of Fable 5.1 — at 64% lower cost, built on a 2.8-trillion-parameter Kimi K3 base with RL post-training.…
A developer analysis of Artificial Analysis benchmarking data finds that GLM-5.3-Flash scores 42 at $0.25 per task while Kimi K3 scores 44 at $2.00 per task, an eight-fold cost gap for a two-point cap…
DeepSeek is retiring its flagship deepseek-v4-pro model, routing all API requests to V4.1-Flash as of September 14, 2026, until V4.1-Pro ships. According to the company, testing by multiple parties fo…
Cognition's SWE-2 coding model scored 83.75% on the eight-task KingBench 3 benchmark, beating DeepSeek V4.1 Flash at 81.25% and its own base model Kimi K3 at 77.5%, according to independent testing. S…
DeepSeek released V4.1 Flash, a 552-billion-parameter mixture-of-experts model with 8 billion active input and 16 billion active output parameters that scored 74.2 on Deep SWE and 30 on Terminal Bench…
DeepSeek launched V4.1-Flash on Thursday, a 552-billion-parameter mixture-of-experts model that activates roughly 8 billion parameters per input token and 16 billion per output token, with a one-milli…
LogRocket's September 2026 AI dev tool power rankings place Anthropic's Claude Fable 5.1 at #1 among 21 models with 1762 Elo on the WebDev Arena, 73 points above Fable 5, while Claude Code took the to…
A technical survey traces five rewrites of transformer attention since 2017, framing every variant as either storing less or reading less. It details how multi-query and grouped-query attention shrink…
Cognition released SWE-2 on September 10, 2026, a coding model it says scores within one benchmark point of Anthropic's Fable 5.1 on FrontierCode 1.1 Main (50.0% vs 50.9%) at 64% lower compute cost, a…
Sakana AI released Fugu Max and Fugu Ultra v2, two new models in its Sakana Fugu family, live today through its OpenAI-compatible API with no open weights and no EU/EEA availability. Fugu Max is price…
Z.ai released GLM-5.3, a 743-billion-parameter model that matches its predecessor GLM-5.2 in size but delivers a 50% gain in programming capability and tops global cybersecurity benchmarks through pos…
A developer extended Sebastian Raschka's GPT-2-style code from the book "Build a Large Language Model (from Scratch)" to add mixture-of-experts support and trained a 446M-parameter model with 220M act…
Cognition released SWE-2, a 2.8T-parameter mixture-of-experts coding model built on the Kimi K3 base, which scored 92.8 on Terminal-Bench 2.1 — the highest figure in the published table — and 50.0 on …
Cognition released SWE-2 on September 10, a coding model it says approaches Fable 5.1's performance at 64% lower mean rollout cost, according to the company's own FrontierCode benchmark. SWE-2 scored …