Qwen

Expert Takeaways: Data Centers 1. Open-Source Models Closing the Gap > Open-source models are catching up to state-of-the-art frontier models faster than expected, driven by lower costs and improved performance. > Starting in Q2, customers…

One thing I also enjoy following is edge AI chips, which are used for edge devices. Axera this week came out w/ AX8860 chip w/ 256 TOPS & 128GB LPDDR5X + 273 GB/s memory bandwidth. Uses 6nm process. This type of chip can run Qwen 27B model…

🎙️ “Hey Luna, take this block. Now give it back. Follow me.” In this demo by @AI_with_Eric, his local AI assistant Luna brings together #LocalAI and #robotics to turn voice commands into physical actions. With the #reBotArmB601, #Qwen 3.8…

Still waiting for like a qwen-3.8 8b or 16b. Im not sure if you will need the vla for pick and place tasks. https://t.co/mFb6wfV6Lk

No tokens. Just a number between 0 and 1. That's the entire output of Liquid AI's d1 family, and this week they shipped two new ones, open weight this time: d1-3B for text and vision, d1-omni-600M for text plus image or audio, both aimed…

China is about to unleash a 60x AI demand shock and it could create an entirely new group of stock market winners. JP Morgan expects China’s inference-token consumption to grow roughly 60-fold between 2026 and 2030. The chart shows demand…

If you're a robotics researcher working with GPT 6 Astra to do interesting things with robotic control or agentic real2sim, and would like access to open source models (Kimi K3, Qwen 3.8 Flash Next) as a comparison, please DM me! I'd love…

Another Chinese lab just gave away a trillion parameter model for free, and the market barely blinked because it's the fourth time this year. Xiaomi open sourced MiMo-V2.6 today: two natively omnimodal models, Pro and Flash, both MIT…

a few notable open model releases *since* this interview of @dylan522p by the way :) Thinking Machines - Inkling - Jul 15 Moonshot - Kimi K3 - Jul 16, weights Jul 26/27 inclusionAI - Ling 3.0 Flash - Jul 23 DeepSeek - V4-Flash-0731 - Jul…

The shift I keep coming back to this week: not every AI call needs to write a sentence. Diogo Almeida's Jev proved that in nine days. Jev is what TypeSafe calls a System One model. It skips token by token generation entirely. Feed it a…

Apple shipped something small this cycle that I think matters more than another point release of a frontier model: macOS 27 now includes an fm command in the terminal, putting Apple's on-device model one shell command away from every…

Open weights image generation just became serious. Qwen-Image-2.1 ships today. Generation and editing in one model. 2K resolution native. Professional typography. Alpha channel. Open weights. The market map: GPT Image delivers impressive…

Open model adoption just hit a tipping point. Here's what that means for builders: 1/5 - The Volume Evidence 78.4% of tokens on Vercel's AI gateway run on open models. That's not a survey. That's production traffic. The switch is already…

Open source just shipped a 744B agentic model. Shanghai AI Lab released Atria Dawn Preview on September 11 with no blog post, no pricing announcement, and no press release. Just a Hugging Face commit and an FP8 checkpoint the next day. The…

A new agentic open-weights model drops Tuesday. 3x cheaper than DeepSeek Flash. 15% better on long agentic loops. Available on RouteLLM API and on the coding agent. Designed for 90% of coding tasks. My read on what's happening in the…

Agentic AI is moving to the edge. On NVIDIA Jetson AGX Thor, TensorRT Edge-LLM completed MLPerf Edge Agentic 6.4x faster than the llama.cpp reference, reaching 52.33 token/sec for Qwen 3.6 27B model. NVFP4, ~96% of prompt tokens served…

🔥🔥RUN #7B #VLM on reComputer #RK3576 Dev Kit + #RK1828 We tested #Qwen 2.5-VL-7B_prune on the reComputer RK3576 Dev Kit + RK1828 AI Accelerator: ⚡ TTFT: 418.93 ms 🚀 Streaming Output: 330.68 chars/s ⏱️ End-to-End: 2.84 s From image…

Jon Cipher is right about the structural shift here. The OpenRouter data tells a clear story. Chinese models took the lead in OpenRouter token share in March and kept it. Not because they win every benchmark. Because restricted compute…

Local inference is becoming a product category, not just a benchmark table. Perplexity launched Portable Computer on Windows RTX PCs today. The orchestrator LLM, subagent LLM, and the full agent harness all run on-device. No cloud…

From Anthropic's new distillation accusations against Chinese AI labs: - Fable wasn't used for distillation. Zai tried but failed. This means Anthropic can stop distillation with sufficiently strong guardrails? Opus was used. - It's not…

🔥🔥RK3576+RK1820: 178.86 tokens/s. Local LLM, Running Fast. We put the #reComputer #RK3576 Dev Kit with #RK1820 AI Accelerator to the test. Running #Qwen 2.5-0.5B, it achieved an impressive 178.86 tokens/s in our test. 🧠 Rockchip RK3576…

I think the point that "astra is the first vlm that doesn't suck" is a really important one. At the smaller scale the qwen models are incredible but you feel the limitations. Still waiting on my qwen 3.8 8b/14b.... https://t.co/Qa2xif9sWG

Open weights are not the same as cheap serving. Alibaba's Qwen 3.8-Max, 2.4T parameters with 95B active, is estimated to need 48-64 B200-class GPUs to serve. DeepSeek's V4 Flash, 284B, fits in about 142GB at FP4. Downloadable and…

A model that optimizes its own training pipeline is a different kind of release. Hy4 is a 770B MoE with 49B active parameters and a 1M-token context, open-sourced by Tencent. The headline number most people cited was the #3 open-source…

The inference cost gap between centralized APIs and decentralized GPU networks is getting hard to ignore. Qwen3.6 35B running at 14,499 tokens per second on a single B200 GPU on Lium. 85% cheaper than the same model on Openrouter. That…

The framing from YC's latest hackathon is different from what I usually see from accelerators: build and own your own intelligence. Not your own product. Not your own LLM wrapper. Specifically: your own agent, your own models, your own…

Qwen3.8-Flash-Next just made the 180B parameter class meaningfully accessible. The architecture is worth understanding: 125B language model parameters, 55B auxiliary, but only 6B activate per token. That sparse activation pattern is what…

Sequence parallelism today partitions the same way regardless of micro-batch size. BASP splits GPUs into disjoint groups sized to the micro-batch, shrinking the all-to-all group. Training time improves 1.17-1.31x for Llama and Qwen, memory…

The hardest thing to benchmark is not intelligence. It is judgment over a full year. E-Commerce Bench, released this week by the Qwen team, is the first benchmark that runs an LLM agent through a full 365-day fiscal year operating multiple…

Zai presentation on AI TAM & long task time horizon was lightbulb for me on where AI is going. Since business pay the bills & they demand Autonomous AI able to operate long period of time, that's where Zai is going. That's also where Qwen…

🔥🔥NEW Launch: reComputer RK3576 Dev Kit/Module/IO Board is Officially HERE!!!! Level up your edge AI development with Rockchip #RK3576 today! 🎉 🧩🧩Dev Kit / Module / IO Board: https://t.co/xQVIT62cbw Powered by #Rockchip #RK3576…

NVIDIA Research 🚀 has produced some great research, like LatentMoE (used in Kimi K3) and GatedDeltaNets (used in Qwen). But for e2e frontier training, NVIDIA's bureaucratic culture has produced embarrassing models like Nemotron3 Ultra.…

Post Log in Sign up Post Chubby♨️ on X: "wtf Qwen 3.8 has been updated to version 0902 and is now almost at the same level as Fable 5 in benchmarks. The interesting two aspects: 1) For such significant jumps, it seems that you don't even…