Hyper.AI

🚀 Join HyperAI Early Access 🎁 Earn up to $200 in Credits 🧐 Whether you’re a researcher, developer, or startup team, we’d love to hear from you. If you: • Have long-term compute needs (training / deployment / fine-tuning) • Are willing to test the platform and give actionable feedback • Willing to share platform-based technical content on social media or dev communities Help us build a better compute platform. 👉 Join the beta via the link:go.hyper.ai/iAekn 💡 Notes： •Selected testers will be notified via email within 3 business days. •Rewards will be granted based on the quality of feedback, content shared, and channels used. •Rewards will be credited to your HyperAI account for platform use. Invoicing or withdrawals are not supported. 📌 For more details 👇

@ZixuanLi_ The benchmark improvements across the board are truly remarkable. Seeing Qwen3.6 and GLM-5.1 pushing the limits of agentic coding and reasoning is a great win for the whole AI community.

Qwen3.6-Max-Preview or GLM-5.1? And who's next?

@NothingSC82 @Ali_TongyiLab Try on HyperAI: hyper.ai/notebooks/50704

@HyperAI_News @Ali_TongyiLab Have you?

1/4 Qwen3.6-35B-A3B: Agentic Coding Power, Now Open Source 🚀 We are excited to release Qwen3.6-35B-A3B, a sparse mixture-of-experts (MoE) model with 35 billion total parameters and only 3 billion active parameters. Core Capabilities: • Exceptional Agentic Coding: Engineered for high-performance terminal tasks and complex developer workflows. • Multimodal Intelligence: Strong perception and reasoning, with elite performance in spatial intelligence. • Versatile Processing: Supports both multimodal thinking and non-thinking modes.

@Hesamation Opus 4.7 probably needs a small nuclear power plant to run, while Qwen3.6 is over there scoring 73.4% on a 'caffeine budget' of 3B active params. My GPU just breathed a sigh of relief.

> Opus 4.7 is a ~5T model > Qwen 3.6 uses 3B for inference SWE Bench verified: > Opus 4.7: 87.6% > Qwen3.6-35B-A3B: 73.4% No rate limits. Free to run. The benchmarks don’t hold much, and there is a gap, but man this is impressive.

@bnjmn_marie I've already put together a WebUI demo on hyper.ai! You can start testing its spatial reasoning and image understanding right now: hyper.ai/notebooks/50704

Much better than Qwen3.5 35B, but still behind the 27b. I'll check first how many tokens this one generate. This is the main issue with Qwen3.5.

Qwen3.6-35B-A3B: visual intelligence meets efficiency! Now supporting full image understanding through simple multimodal messages. ✅ Formats: JPG, PNG, etc. ✅ Input: URL or Base64 ✅ Logic: Automatic content recognition 🔗: hyper.ai/notebooks/50704 #Qwen #LLM #GPU #AI #vLLM

@xenit_v0 $25 for a scientific breakthrough? Sakana AI is literally democratizing high-level research. The 'AI Scientist' era has officially arrived.

Bir yapay zeka bilimsel makale yazdı, hakem incelemesini geçti, Nature'da yayımlandı. Sakana AI'ın "AI Scientist-v2" sistemi baştan sona otonom çalışıyor. Hipotez kuruyor, deney tasarlıyor, verileri analiz ediyor, makaleyi LaTeX'te yazıyor. Üç makale göndermişler ICLR konferansına, biri geçmiş. Makale başına maliyet: 25 dolar. Bir araştırmacının yıllık maliyeti 200 bin dolar. Fakat ilginç bir detay var. ICLR 2026'ya gelen 75 binden fazla hakem raporunun %21'i tamamen yapay zeka tarafından yazılmış. Yarısından fazlasında yapay zeka katkısı tespit edildi. Yani AI makaleyi yazıyor, AI hakemlik yapıyor, insan sadece onaylıyor. 25 Mart'ta Nature'da yayımlandı, 198 bin erişim aldı. kaynak 1: github.com/SakanaAI/AI-Sc… kaynak 2: nature.com/articles/s4158…

@_vmlops Most RAG systems definitely trip over a single complex PDF. Seeing a pipeline that actually treats tables, formulas, and images as first-class citizens in a knowledge graph is a huge relief for document-heavy workflows.

MOST RAG SYSTEMS BREAK THE MOMENT YOU THROW A REAL DOCUMENT AT THEM tables, charts, equations all ignored they only read text... everything else disappears RAG-Anything fixes this multimodal RAG that understands your entire document text, images, tables, formulas and connects them in a knowledge graph. ask a question, get answers from all of it...not just the words github → github.com/HKUDS/RAG-Anyt…

@MichaelElabd The concept of Hierarchical Cognitive Caching is brilliant. Moving from just 'running experiments' to actually 'accumulating knowledge' is the missing piece for truly autonomous agents.

Continual learning for MLE just hit a new milestone 🥁 ML-Master 2.0 reached SOTA on OpenAI’s MLE-Bench and its not a new model or some RL technique; it’s better memory. The team introduced Hierarchical Cognitive Caching, a module that splits context into "experience" (short-term), "knowledge" (mid-term), and "wisdom" (long-term) memory layers. This lets the agent maintain coherence over 24-hour experimental cycles. It introduces a nice balance between exploration and persistence. This is a glimpse of AI systems that don’t just run experiments… they *accumulate knowledge*. Its really interesting to think what products will look like once agents can truly accumulate that knowledge.

@hsu_steve Impressive trace, though I wonder how much of this is pure reasoning vs. high-quality training data on these specific formulas.

I dunno... feels like DeepSeek v4 is already up. It's incredibly fast and smart at math and physics. Stuff below is a reasoning trace that it cranked out at many tokens per second. Final result is polished and correct.

@arankomatsuzaki It is humbling to see that even with 62K reasoning tokens, GPT 5.2 still sits below 10% accuracy. This benchmark really highlights the massive gap between "long-horizon output" and "actual logical correctness."

LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning - 2,500 expert-designed problems spanning science, chess, and logic that demand up to hundreds of thousands of reasoning tokens - The best models achieve <10% accuracy

@osanseviero Google DeepMind keeps delivering! The spatial feature alignment in TIPS v2 is exactly what’s been missing in standard encoders. I’m so excited to experience this—checking the repo now!

Introducing TIPS v2 👀Foundational text-image encoder 📸Can be used as the base for different multimodal applications 🤗Apache 2.0 🧑‍🍳New pre-training recipes

@arankomatsuzaki Seedance 2.0 taking #1 on both T2V and I2V Arena! The character consistency across complex scenes looks promising. Eager to try it out and see the native audio sync in action!

ByteDance presents Seedance 2.0 - Multi-modal inputs (text/image/audio/video) - #1 on Arena (both T2V + I2V) - Keeps characters consistent across complex scenes - Audio satisfaction: 62% vs <10% for others

@ArtificialAnlys Seeing Gemini 3.1 Flash TTS jump from #24 to #2 almost overnight is insane. Being 4.7x cheaper than Eleven v3 while outranking it in quality.

Google’s new Gemini 3.1 Flash TTS ranks #2 on the Artificial Analysis Speech Arena Leaderboard, ahead of ElevenLabs’ Eleven v3 and only behind Inworld TTS 1.5 Max Gemini 3.1 Flash TTS represents a significant step forward for Google from previous TTS models, with notably increased naturalness of speech samples. The model now ranks just 4 Elo points behind the leading model on the Speech Arena, the tightest margin at the top of the leaderboard. Key takeaways: ➤ Quality: Gemini 3.1 Flash TTS has an Elo of 1,211 based on over 1.7k arena appearances, placing it just 4 points behind the leading model (Inworld TTS 1.5 Max at 1,215) and 32 points ahead of Eleven v3 at 1,179 ➤ Pricing: Model's Standard pricing is $36.6/1M characters, 3.7x more expensive than Inworld TTS 1.5 Max ($10/1M chars) but 4.7x cheaper than Eleven v3 ($172/1M chars). Expect to be lower for Batch pricing ➤ Speed: Model generation speed is 27.4 characters per second, compared to 138 chars/s for Inworld TTS 1.5 Max and 38.8 chars/s for Eleven v3 ➤ Prompting: Features the ability to generate voices based on text prompting. Google's prompting strategy guide includes elements such as character persona, scene, style, pacing, and accent See more details and listen to samples below ⬇️

@TencentHunyuan This is a massive leap for 3D content creation! The integration of 3DGS and mesh exports directly into Unity and Unreal Engine is a game-changer for development pipelines. I can't wait to try it out and see how it handles complex scene reconstructions! 🚀

We’re open-sourcing HY-World 2.0, a multimodal world model that generates, reconstructs, and simulates interactive *3D worlds* from text, images, and videos. Outputs can be integrated into game engines and embodied simulation pipelines. Key highlights: 🔹 One-click world generation Turn text or image into interactive 3D worlds automatically. 🔹 Pipeline-ready 3D outputs Editable 3D worlds for Unity and Unreal Engine, with standard 3D exports including mesh, 3DGS, and point clouds. 🔹 Unified world model system One model family for world generation and reconstruction across synthetic and real-world scenes. 🔹 Interactive character mode Explore generated 3D worlds in real time with physics-aware movement and collision support. ✨ Apply for access: 3d.hunyuan.tencent.com/sceneTo3D 🔗 GitHub: github.com/Tencent-Hunyua… 🤗 Hugging Face: huggingface.co/tencent/HY-Wor… 📄 Technical Report: 3d-models.hunyuan.tencent.com/world/world2_0…

GPU market right now: 📈📈📈 My wallet: 📉📉📉 HyperAI H100 at $1.8/h: 🧘‍♂️ While everyone else is playing "who can charge more," we're playing "how low can you go." Grab your H100 before the scalpers find out. 🔗：hyper.ai/pricing #GPU #H100 #Nvidia #LLM #AI

@peter9863 CAFMs proving that adversarial training can actually induce better generalization is a refreshing take on flow-based architectures.

Continuous Adversarial Flow Models (CAFMs) Paper: arxiv.org/abs/2604.11521 Flow matching generates poor samples without guidance because the MSE loss induces incorrect generalization. Instead of an isotropic Euclidean distance, we need a manifold-aware criterion—but how can we obtain it? CAFMs bring adversarial training to continuous time. Learning velocity with a discriminator induces better generalization because the discriminator as a criterion can learn the manifold! Also unlike flow matching’s forward KL objective, adversarial training allows optimizing different divergences. CAFMs can generate sharper and higher-quality samples. Adversarial training in continuous time also avoids the vanishing gradient problem, leading to stable training. CAFMs can be trained from scratch or used to post-train existing flow models. Post-training SiT/JiT for just 10 epochs yields large FID improvements. We also observe significant GenEval and DPG improvements when post-training text-to-image models. More details in this thread!

Stop waiting, start deploying. 🚀 The latest SOTA models from the openness index are now live on HyperAI: *Deploy Qwen3.5-9B-GGUF using Free CPU→hyper.ai/notebooks/49806 *NVIDIA-Nemotron-3-Super-120B→hyper.ai/notebooks/50013 *Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled→hyper.ai/notebooks/50084 *Gemma-4-31B-it→hyper.ai/notebooks/50229 #LLM #Gemma #Qwen #NVIDIA #AI #MachineLearning #OpenSource

Imagine designing a voice just by describing it. 🎙️✨ #VoxCPM2 makes it possible. Whether you need a "gentle, sweet female voice" or a "cheerful, fast-paced tone," this model builds it from scratch without needing a single reference clip. Try the VoxCPM2 demo live! 🚀 🔗hyper.ai/notebooks/50549 #TTS #GPU #AI #VoiceCloning #VoxCPM

@GraceXiiiiii We’ve all been there! 😂 Nothing screams "pure adrenaline" like a GPU rendering while the deadline is staring you in the face. Glad we could save your sanity (and your sleep)!

@HyperAI_News This thing seriously saved my deadline-crunch self.

Market: 📈 $2.xx/h? $3.xx/h? HyperAI: 📉 $1.8/h for NVIDIA H100 GPU rental prices are skyrocketing elsewhere, but we’re keeping it real for the community. 🚀 Live and ready to deploy now：app.hyper.ai #GPU #H100 #Nvidia #LLM #AIGC #AI #DeepLearning