Loading...
9 entries with this tag
On September 18, 2026, Alibaba Qwen released Qwen3.8-Omni-Flash — a native omnimodal model that accepts text, image, audio, and video inputs with a 1M-token context window, adjustable reasoning, function calling, and web search. At $0.15/$0.47 per million tokens with $0.016 cached input, it reports 25%+ average improvement across 29 evaluations over Qwen3.5-Omni-Plus, with particular breakthroughs in multi-speaker meeting analysis (82.7 on LongAudioSpan, 3.48% DER on AliMeeting) and agentic video benchmarks (71.0 on WildClawBench-MM, a 106% jump). The model is built on the Qwen3.8-Flash-Next architecture and ships with Qwen-MM-Plugins, an Apache-2.0 toolkit for multimodal agent harnesses.
Two new research articles published: AI News Weekly covering the record week of GPT-6 Sol/Luna, Claude Opus 5.5, MiMo-V2.6-Pro, US-China AI hotline, and Trump's AI Force; plus a deep-dive on MiniMax M3.1-Flash-Preview's stealth launch, six-tier effort ladder, and forced reasoning.
On September 18-20, 2026, StepFun released Step 5 Preview — a 600B-parameter sparse MoE model with only 27B activated per token, achieving 67.7% on DeepSWE v1.1, 33.3% on Terminal-Bench 4.0, and 99.8 tokens/sec output speed at $1/M input pricing, with open weights scheduled for October 15, 2026.
On August 21, 2026, DeepSeek launched V4-Flash-Vision-Exp on its API, followed by open weights on Hugging Face under MIT license on August 31. This is DeepSeek's first multimodal model, adding a vision encoder and aligner to the V4-Flash backbone (284B total, 13B active). It achieves 36.5% on ApexBench (vs. 26.2% for text-only V4-Flash), 27.3% on Agents' Last Exam (beating Opus-4.8's 25.7%), and 35.0% on ZeroBench Pass@5. The MIT license — without the commercial restrictions seen in Qwen's Community License — makes it the most permissively licensed frontier multimodal model available.
On August 26, 2026, Alibaba's Qwen team released Qwen3.8-Flash-Next — a 125B total / 6B active multimodal MoE that serves as the first open-weight preview of the Qwen4 architecture. Introduces N-gram Embedding (51B params), Qwen Sparse Attention (QSA), Gated Residual, and Muon+AdamW hybrid optimization. Covers architecture, benchmarks vs. DeepSeek V4-Flash and Claude Opus 4.6, deployment guidance, and the Qwen4 roadmap.
One new research article published: comprehensive deep-dive on Moonshot's Kimi K3 — the world's first open 3T-class model at 2.8T parameters with 104B active, built on Kimi Delta Attention and Attention Residuals. Covers architecture, training, benchmarks vs. Fable 5 and GPT-5.6 Sol, and the open frontier trajectory.
On July 16, 2026, Moonshot AI announced Kimi K3 — the world's first open 3T-class model at 2.8T parameters with 104B active, built on Kimi Delta Attention, Attention Residuals, and Stable LatentMoE (16/896 experts). Weights released July 27 under Kimi K3 License. Covers architecture, training, benchmarks vs. Claude Fable 5 and GPT-5.6 Sol, pricing, and the open frontier trajectory.
Google DeepMind releases Gemma 4 12B — a 12B dense model with encoder-free multimodal architecture, native audio support, and 256K context. Runs on 16GB laptops under Apache 2.0. Benchmarks approach the 26B MoE sibling at less than half the memory. The most practical multimodal model for local deployment yet.
May 4: AI News Weekly published covering frontier cyber-offense capabilities crossing a critical threshold (Claude Mythos & GPT-5.5 clearing 32-step simulations), Chinese open-weights models narrowing competitiveness gap, mega-rounds reshaping lab economics ($122B OpenAI, $45B+ Anthropic), and dual-use policy tensions escalating. Key insight: Frontier labs now operating in parallel channels—capability announcements coordinated with security institute reviews; infrastructure consolidation accelerating via mega-rounds + Chinese open-weights competition.