Hy4·Wan3.0·Qwen3.8-Flash 확산과 저비용 AI 연구·코딩 에이전트 운영

Hy4 preview is here. 770B, 49B active, 1M context. Built for productivity. Open source frontier. Consistent affordable price. Use it. Tell us what breaks. More on Hy blog： https:// hy.tencent.ai/research/hy4-p review … HuggingFace： https:// huggingface.co/tencent/Hy4-pr eview … Github： https:// github.com/Tencent-Hunyua n/Hy4-preview …

@TencentHunyuan 's Hy4-preview runs in vLLM from day 0, verified on @NVIDIA Blackwell. - 770B total, 49B active, 256 routed experts plus one shared - 1M context, but each query attends to just 2048 tokens - Only 21 of the 78 layers compute their own sparse index, the other 57 reuse one - A 10B MTP layer ships inside the checkpoint, 0.7B of it active, draft depth 3 Tencent's HPC-Ops attention and MoE kernels have been in vLLM main since Hy3. Hy4 brings them to Blackwell. VLLM_ENABLE_HPC_OPS=1 vllm serve tencent/Hy4-preview-FP8 -tp 8 Thanks @TencentHunyuan for the preview weights! http:// recipes.vllm.ai/tencent/Hy4-pr eview …

Wan 3.0 is now live on @Medeo_AI ! Native 30-second generation, synchronized audio, and strong character & style consistency. Build complete cinematic scenes on Medeo. More API access： • Model Studio: https:// click.alibabacloud.com/m/20000002017/ • Qwen Cloud: https:// click.qwencloud.com/m/20000002025/

Wan 3.0 is now on @DeepInfra ! Video + audio in one model, with 30-second generation and multimodal references. Now available on DeepInfra. More API access： • Model Studio: https:// click.alibabacloud.com/m/20000002017/ • Qwen Cloud: https:// click.qwencloud.com/m/20000002025/

OpenCode

@opencode

1일 전

Qwen3.8-Flash is now available in OpenCode Go 125B/6B · 1M context · multimodal

Qwen3.8-Flash is now available on @gmi_cloud — a multimodal MoE and an early preview of the Qwen4 architecture. More API access： • Model Studio: https:// click.alibabacloud.com/m/20000002820/ • Qwen Cloud: https:// click.qwencloud.com/m/20000002828/ #BuildWithQwen

We’re gonna start testing our first V8.2 edit model today. This model supports editing with instructions, generating images with other images (up to 4 references), brush-based inpainting, and outpainting. It also works with personalization, moodboards, and srefs. Have fun!

Accelerating Scientific Research with Gemini in the Real-World GDM applied Co-Scientist to three different domains: 1. Materials Science: interfaced with a semi-automated CVD instrument to design a non-hazardous precursor route (C2Cl6) targeted for Ti3C2T𝑥 MXene growth. Physical execution of the top-ranked recipe optimized by human experts yielded 2D layered structures exhibiting similar characteristics to Ti3C2T𝑥 MXene 2. Biology: Co-Scientist built a system to predict emergent swarming phenotypes of engineered E. coli across inducer (IPTG) concentration gradients from sparse imaging data. These predictions were quantitatively validated against unpublished wet-lab morphological measurements 3. Computer Science: Co-Scientist operated fully autonomously to discover an inference-time scaling architecture that outperformed six frontier models on HealthBench Hard and Professional while achieving a significant, though modest, reduction in potential clinical harm under blinded physician evaluation "Spanning ideation, experimentation, and manuscript generation, Co-Scientist dynamically adapts the level of human–AI collaboration to the physical constraints and verification requirements of each scientific domain" link: https:// arxiv.org/abs/2608.26701

Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090 "Our best checkpoint surpasses Qwen2-1.5B overall and approaches Qwen2.5-1.5B." "These stages total 1.4T scheduled tokens and 22,514 active-training GPU-hours, corresponding to a compute cost of about $6.9K and 17.6 elapsed days. The pipeline covers the data recipe, hardware infrastructure, supporting software, and training strategy." link: https:// arxiv.org/abs/2608.27370

LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics "We introduce LeVJEPA, the first video encoder trained under LeJEPA's collapse-free objective" "At matched epochs on identical data, LeVJEPA matches or surpasses V-JEPA 2 across ViT-S/B/L at 5.6 to 20.8x less pretraining compute, and at matched total FLOPs it exceeds the strongest video baseline by 7.6 points on ImageNet-1K while remaining competitive on motion-centric benchmarks." "Against a compute-matched DINOv2 trained on frames of the same videos, LeVJEPA approaches the image-pretrained encoder on appearance-centric evaluation while nearly doubling its motion-centric accuracy." link: https:// arxiv.org/abs/2608.27395

Uber: more than 70% of pull requests are now attributed to local or cloud agents. But the more interesting part is how they’re making agentic coding economically sustainable: 3,600+ skills, 30K+ executions/day, and a sharply lower cost per session through better context management, tool loading, and orchestration. The takeaway: the competitive advantage is no longer just the model. It’s the whole agent harness.