Latest AI Technology

Track the latest breakthroughs and technological advances in AI

69 articles

Alibaba Qwen 3.8 Max Deep Dive: 2.4 Trillion Parameter MoE Architecture, 1M Token Context, Autonomous 10-Day Software Development, Breaks into Global Top 10 Model Rankings

Alibaba released Qwen 3.8 Max on August 3, 2026 — a flagship MoE model with 2.4 trillion total parameters, 95 billion active per query, 1M token context, and multimodal capabilities. It ranks #10 of 232 models on BenchLM and #5 in multimodal tasks.

Sep 17, 20269

Shanghai AI Lab Quietly Releases Atria Dawn Preview: 744B Parameter Open-Weight Agentic MoE Model, MIT License, BrowseComp 92.5, Challenging Closed-Source Frontier Models

Shanghai AI Lab quietly released Atria Dawn Preview on September 11, 2026 — a 744B parameter open-weight agentic MoE model with MIT license, scoring 92.5 on BrowseComp and 96.0 on DeepSearchQA, using a 'Verifiable Experience Pipeline' training paradigm to challenge closed-source frontier models, though performance data awaits third-party verification.

Sep 16, 202610

OpenAI GPT-6 Astra Officially Released: First to Hit 'Critical' Cybersecurity Threshold, 99.9% ARC-AGI-3, Ushering in New Era of Agentic AI

OpenAI officially released GPT-6 Astra on September 3, 2026, becoming the first model to trigger the company's 'Critical' cybersecurity threshold, achieving 99.9% on ARC-AGI-3 and 97.6% on FrontierMath Tier 4, with API pricing at $10/M input and $50/M output tokens, plus a 1M-token context window and cross-context memory for Codex.

Sep 14, 20269

Google DeepMind WeatherNext 3 Officially Released: Hourly Updates, 5km Resolution, 50% More Accurate Precipitation Forecasts, Integrated into Search, Maps and Gemini

Google DeepMind released WeatherNext 3 on September 3, 2026, using a Functional Generative Network with mesh transformer architecture for hourly 15-day global probabilistic forecasts. Precipitation accuracy improved by up to 50%, now integrated into Google Search, Maps, and Gemini.

Sep 13, 202611

Sakana Fugu Ultra v2 Officially Released: Multi-Agent Orchestration Achieves 74.3% DeepSWE, $30 Per Million Output Tokens, Challenging Single Frontier Model Dominance

Sakana AI released Fugu Ultra v2 on September 11, 2026. This multi-agent orchestration model, grounded in ICLR 2026 TRINITY and Conductor research, achieves best or joint-best scores on five internal benchmarks, with 74.3% on DeepSWE and 48.3% on Chartography. The model deliberately excludes Claude Fable 5 and GPT-6 Astra to prove orchestration architecture capability, priced at $5 per million input tokens and $30 per million output tokens.

Sep 12, 202611

DeepSeek V4.1 Flash Officially Released: 552B MoE Causal Encoder-Decoder Architecture, Sub-$0.01 Per Million Tokens, Asia-Pacific Markets Rattled

DeepSeek officially released V4.1 Flash on September 10, 2026, featuring a novel Causal Encoder-Decoder architecture with 552B MoE parameters activating only 8-16B per token, priced below $0.01 per million tokens, triggering significant volatility in Asia-Pacific tech stocks.

Sep 11, 20269

OpenAI GPT-6 Astra Officially Released: First 'Critical' Cybersecurity-Rated Model with 1M Token Context and Recurrent Depth Reasoning

OpenAI officially released GPT-6 Astra on September 3, 2026, becoming the first AI model to reach a 'Critical' cybersecurity rating, scoring 97.6% on FrontierMath Tier 4 and 100% on ExploitBench, featuring recurrent depth reasoning architecture at $10/M input and $50/M output tokens.

Sep 9, 20268

Google Releases Gemini 3.8 Flash: Third Flash Update in Six Weeks, Restricted Cyber Security Variant Debuts

Google released Gemini 3.8 Flash on September 2, 2026—the third Flash series update in six weeks. The model excels in agentic workflows, long-horizon coding, and complex reasoning, maintaining pricing at $0.75 per million input tokens. The simultaneously launched Gemini 3.8 Flash Cyber security variant delivers 2.6x patch accuracy improvement but is restricted to trusted defenders only.

Sep 8, 20267

Firmus Closes $2B Funding Round: Valuation Surpasses $10.5B, NVIDIA Partnership Powers 1.6GW Australian AI Factory, Asia-Pacific Expansion Accelerates

Australian AI infrastructure company Firmus closed a $2 billion strategic equity round in August 2026, with valuation surpassing $10.5 billion — nearly doubling from April. NVIDIA, Coatue, Blackstone, and Jane Street participated, with funds accelerating Project Southgate's Australian AI factory plan targeting 1.6GW of compute capacity by 2028, and expansion into Indonesia and other Asia-Pacific markets.

Sep 7, 20267

Meta Muse Spark 1.3 Officially Released: 1M Token Context Window, 20% Fewer Tool Calls, Multimodal Agentic Model Challenges Claude and GPT

Meta officially releases Muse Spark 1.3 with a 1M token context window, scoring 98.1 on the 512K-1M MRCR benchmark. Engineering workflows see 20% fewer tool calls and 25% fewer tokens. Priced at $1.25/$4.25 per M tokens, currently available to developers via Meta Model API only.

Sep 5, 20267

OpenAI GPT-6 Astra Officially Released: 99.9% ARC-AGI-3 Score, Perfect ExploitBench, First AI Model to Reach 'Critical' Cybersecurity Threshold

OpenAI officially released GPT-6 Astra on September 3, 2026, achieving a 99.9% ARC-AGI-3 score and a perfect 100% on ExploitBench, making it the first AI model to reach the 'Critical' cybersecurity threshold. Trained on 100,000 GPUs, the model uses a staged rollout with offensive cyber capabilities restricted to vetted defenders in the Daybreak Blue program.

Sep 4, 202612

Google Gemini 3.8 Flash Officially Released: Three Updates in Six Weeks Sets Record, Flash Cyber Security Variant Improves Patch Accuracy by 2.6x

Google releases Gemini 3.8 Flash and security-specialized Flash Cyber, completing three Flash series updates in six weeks. 3.8 Flash tops the DeepSWE v1.1 leaderboard, while Flash Cyber improves patch accuracy by 2.6x over larger commercial models, restricted to trusted testers and government entities.

Sep 3, 202611

PreviousPage 5 / 6Next