Gemini 3.6 Flash: Google's New Agent Workhorse
Google shipped Gemini 3.6 Flash on July 21, 2026: $7.50 output pricing (down 17%), 17% fewer output tokens, and stronger agentic coding. What it means for builders.
Mozilla State of Open Source AI 2026: Key Findings
Mozilla released its inaugural State of Open Source AI report on July 14, 2026. Open-weight models have closed the capability gap with closed frontier systems to just 3.3 percent, while inference costs fell roughly 50 times in three years.
SWE-1.7: Cognition's Coding Model Nears Frontier
Cognition SWE-1.7 lands within a few points of GPT-5.5 and Claude Opus 4.8 on coding benchmarks, at a fraction of the cost. It is live in Devin at 1000 tokens per second.
LongCat-2.0: China's 1.6T Open-Weights Coding Model
Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter agentic coding model trained entirely on Chinese chips. The weights just landed on Hugging Face, and it beats GPT-5.5 on SWE-bench Pro.
Beyond LoRA: Hugging Face Tests Better Fine-Tuning Methods
Hugging Face benchmarked LoRA against newer PEFT methods like OFT, which beat it on image generation while using less memory. The default is not always the best fit.
Google Builds a Skills Marketplace for Gemini
Google is building a Skills Marketplace for Gemini Enterprise, a hub where teams pick predefined AI skills instead of configuring each capability by hand.
Anthropic Disables Fable 5 and Mythos 5 on US Order
Anthropic disabled Claude Fable 5 and Mythos 5 for every customer worldwide after a US government export-control directive on June 12, 2026.
Kimi K2.7-Code: Open Model Beats Opus 4.8 on Tool Use
Moonshot released Kimi K2.7-Code on June 12, an open-weights coding model that beats Claude Opus 4.8 on MCPMark tool use at far lower cost.
Anthropic Apologizes for Fable 5's Invisible Guardrail
Anthropic apologized on June 11 for a hidden Claude Fable 5 guardrail that quietly degraded results instead of refusing them, and said it will make the safeguard visible.
Apple WWDC 2026: New AI Models, Image Playground, Siri AI
Apple WWDC 2026 unveiled five new Foundation Models, a rebuilt Siri AI, Image Playground, and Spatial Reframing, built with Google Gemini collaboration.
Xiaomi MiMo Hits 1000 Tokens Per Second on 1T Open Model
Xiaomi MiMo-V2.5-Pro-UltraSpeed claims 1000 tokens per second decode on a 1T MoE. Open-weights FP4 checkpoint plus a 2-week free API trial.
MAI-Code-1-Flash: Microsoft's New Coding Model in Copilot
Microsoft used its Build 2026 keynote on June 2 to ship MAI-Code-1-Flash, an in-house coding model that is live today inside the GitHub Copilot model picker for Free, Pro, Pro+, and Max tiers.
Microsoft MAI-Thinking-1 Tops Claude Sonnet in User Evals
Microsoft's MAI-Thinking-1, a 35B sparse MoE with 256k context, tops Claude Sonnet 4.6 in blind evaluations and matches Claude Opus 4.6 on SWE-Bench Pro.
Nvidia Adopts OpenMDW for Cosmos and Nemotron Open Models
Nvidia is bringing Cosmos, Nemotron, GR00T, and Ising under OpenMDW-1.1. Here is what the unified AI model license means for creative AI developers.
Xiaomi MiMo-v2.5 Cuts API Prices 99% Across Multimodal
Xiaomi cut MiMo-v2.5 API prices up to 99% effective May 27 across chat, image, audio, video understanding, and TTS, with OpenAI and Anthropic compatible endpoints.
Bonsai Image 4B: 1-Bit AI Image Gen Runs on iPhone
PrismML released Bonsai Image 4B with 1-bit and ternary checkpoints under Apache 2.0. The model retains 95% of FLUX.2 Klein 4B quality at 6.4x smaller size and runs directly on iPhone.
Qwen3.6-27B: Open Dense Model Beats Its 35B Sibling
Alibaba released Qwen3.6-27B on April 22, 2026. The 27B dense open-weight model beats its larger 35B-A3B sibling on coding, agent, and vision benchmarks.
Qwen3.6-Max-Preview: Alibaba's New AI Tops Coding Benchmarks
Alibaba released Qwen3.6-Max-Preview on April 20, 2026, topping six programming benchmarks including SWE-benchPro. Here is what it means for creators building AI workflows.
Meta Muse Spark Marks Proprietary Shift for AI
Meta launched Muse Spark on April 8, the first AI model built by its new Superintelligence Labs division. Unlike every Llama release before it, Muse Spark is proprietary.
Mystery Model HappyHorse-1.0 Tops AI Video Leaderboard
A model called HappyHorse-1.0 has taken the top spot on Artificial Analysis text-to-video leaderboard with an ELO rating of 1365, beating Seedance 2.0 and Kling 3.0 Pro.
Cognition SWE-1.6 Ships Free Coding Agent at 950 Tokens/Sec
Cognition released SWE-1.6, making it generally available in Windsurf with a free tier for three months. The model improves SWE-Bench Pro scores by over 10% at up to 950 tokens per second.
GLM-5.1 Ships Open-Weight 754B Model That Tops SWE-Bench Pro
Z.AI released GLM-5.1, a 754-billion-parameter open-weight model that claims the top spot on SWE-Bench Pro with a score of 58.4, surpassing Claude Opus 4.6 and GPT-5.4.
Meta Plans Open-Source Mango Multimedia AI Generator
Meta confirmed it will release open-source versions of its upcoming Mango multimedia generator and Avocado LLM, extending its open-weight strategy to next-gen frontier models.
DeepSeek V4 Will Run Exclusively on Huawei Chips
DeepSeek V4 will run exclusively on Huawei Ascend 950PR processors, marking a milestone in China's push to build advanced AI without American chip technology.