Alibaba Plans Qwen Model Reaching 10 Trillion Parameters
Alibaba says future Qwen models could reach 5–10 trillion parameters as it links model scaling to domestic chips, cloud capacity and agent deployment.
September 21, 2026
All briefings and quick hits published that day.
- ⚡ Quick hits
- AIGC Tripo announced an additional $10,000 prize pool for its Tripothon S1 challenge with Jupiter SR hardware and software. X @tripoai ↗
- Industry NVIDIA introduced Personal AI Router to assign AI tasks across local computing resources. InfoQ 中文 ↗
- AIGC Tripo showcased a fully playable dark-fantasy game built in seven hours using Tripo3D assets, Fable 5.1, and Godot. X @tripoai ↗
- Agents Kuaishou described a Harness Loop for continuously improving its e-commerce recommendation agent. InfoQ 中文 ↗
- Research Anthropic is setting up a biology lab where Claude guides robots through drug experiments. The Decoder ↗
- Agents Tencent Hunyuan introduced WebCraftBench to test whether agents can build functional websites rather than merely claim completion. X @TencentHunyuan ↗
- AIGC The AI MediaKit team introduced an entertainment production and management platform called WhaleSharp AI. 量子位 ↗
- Industry Ollama said some DeepSeek-V4.1-Flash requests were incorrectly charged, consuming more usage than expected. X @ollama ↗
NVIDIA Open-Sources SoL-Pi for Leaner Coding Agents
NVIDIA’s SoL-Pi extension uses automated research loops to reduce coding-agent token traffic by up to 49% while preserving task performance.
- ⚡ Quick hits
- Agents Kimi K3 became available on Amazon Bedrock for coding, document analysis, and extended agent workflows. X @Kimi_Moonshot ↗
- Agents OpenAI highlighted a surgeon using Codex to build personal tools and review scientific literature on PubMed. X @OpenAIDevs ↗
Tencent Releases Hy Image3.5 Preview for Iterative Creation
Tencent’s Hy Image3.5 preview adds multi-turn image editing and reference inputs, targeting production workflows rather than one-shot generation.
- ⚡ Quick hits
- Benchmark Artificial Analysis found Gemini 3.1 Flash TTS leading its benchmark at 88.1%, while SpaceXAI TTS offered a lower price above 80%. X @ArtificialAnlys ↗
- AIGC Pika launched Relight Media, which rebuilds lighting and shadows in photos or videos after capture. X @pika_labs ↗
- Models Fireworks evaluated 18 models across 113 coding tasks to study routing each task to its strongest model. X @FireworksAI_HQ ↗
- Open Source Hugging Face made Transformers run llama.cpp quantized models, simplifying access to compressed local checkpoints. Hugging Face ↗
- Open Source Hugging Face hired oMLX creator Jun Kim to support the MLX community. Hugging Face ↗
- Industry Waymo launched a San Francisco Bay Area program that rewards riders with $2.85 in Waymo Cash when they combine rides with transit. Waymo ↗
Xiaomi Open-Sources MiMo-V2.6 After Large-Scale RL
Xiaomi releases MiMo-V2.6 Pro and Flash with open weights, RL environments and training tools for agent research.
- ⚡ Quick hits
- AIGC Runway expanded Workflows with compositing, alpha, HDR, depth-map, and RGB-depth tools for more complete creative pipelines. X / Runway ↗
- Agents GitHub said one engineer and a team of agents ported its Copilot agent runtime to Rust, shipping 800,000 lines of production code. X / GitHub ↗
- Industry AWS announced that xAI’s Grok 4.6 is available through Amazon Bedrock for managed enterprise model access. AWS Machine Learning ↗
- Research HeyGen introduced Code2Video, a benchmark developed with Google DeepMind to measure whether generated code renders into watchable video. X / HeyGen ↗
- Agents LlamaIndex added Grounded Confidence to Extract so agents can estimate the reliability of extracted data in workflows. X / LlamaIndex ↗
xAI Releases Grok 4.7 for Coding and Knowledge Work
xAI says Grok 4.7 improves long-running coding and professional tasks while launching at $2 per million input tokens.
- ⚡ Quick hits
- AIGC ElevenLabs released Studio 4.0 with multi-track editing, comments, model access, and an AI co-editor called Studio Agent. X / ElevenLabs ↗
- Agents Krea made its Agent available through MCP, enabling agents to generate assets and automate video pipelines inside workflows. X / Krea ↗
- AIGC Perplexity Computer can generate videos using MiniMax H3 and ByteDance Seedance 2.5, according to Perplexity’s announcement. X / Perplexity AI ↗
- AIGC ByteDance launched Dramagic, a platform designed to take short dramas from scripts through production with AI. The Decoder ↗
AWS Releases Open-Source Strands Agent Harness
AWS launches Strands Harness as a ready-made open-source agent stack, claiming lower token use at comparable task accuracy.
- ⚡ Quick hits
- Research Microsoft Research open-sourced RetroChimera, a retrosynthesis model whose routes received 90% expert acceptance in tests. Microsoft Research ↗
- AIGC Hyper3D introduced Agentic Mode for editable N-GON models, natural-language revisions, animation presets, and multi-view reconstruction. X / Hyper3D ↗
- Funding SoftBank reportedly plans to borrow more than $11 billion through bonds to finance its investment in OpenAI. The Decoder ↗
- Research Bristol researchers argued that medicine’s established methods for handling black-box decisions could inform explainable AI design. The Decoder ↗
- Open Source Hugging Face published a method that treats LLM block pruning as an Ising optimization problem to reduce model size. Hugging Face ↗
OpenAI Forms Independent Mathematics Advisory Group
OpenAI creates an external mathematician group to advise on sharing AI-generated results and responsible mathematical research.