Training utilized 50,000 Chinese ASICs and zero GPUs
WHOA Ok, here it goes. Owl Alpha. V4-scale model, updated DSA, N-gram Embeddings, decent evals, «pre-trained on over 50K AI ASICs … superpods» from *checks notes* Meituan. Yeeeah. these are clearly *not* Ascend 950DT SuperPODs (<<80 GB memory/ASIC). What the hell did they use?!
@teortaxesTex You missed the point. LongCat 2.0 is trained on domestic accelerators
https://longcat.chat/blog/longcat-2.0/
Positive users praise LongCat-2.0's agentic benchmark wins and training on 50K Chinese chips as amazing progress, while negative users dismiss it as benchmaxed garbage or criticize the surrounding hype.
No Digg Deeper questions have been answered for this story yet.
how the fuck did they train a 1.6t model for 35t tokens on asics
Most popular model on @OpenRouter (10tr tokens) turns out to be a 1.6tr MoE by @Meituan_LongCat (superapp/DoorDash of China)
Basically Gemini / Opus 4.6 level
35tr tokens trained entirely on 50k Chinese ASICs
No GPUs needed
https://longcat.chat/blog/longcat-2.0/
Some of you guessed right. 👀 Owl Alpha on @OpenRouter — that's us.
Since going live, it has reached Top 3 globally by daily volume — and #1 on Hermes Agent, #2 on Claude Code, #3 on OpenClaw by monthly volume.
Thank you to everyone who tested and used Owl Alpha during stealth — you helped shape what's coming next. Owl Alpha will be retiring soon. But this isn't an ending — stay tuned!
Meituan, basically China’s DoorDash, trained a 1.6T parameter LLM on 50K Chinese chips.
It reminds me of Jensen Huang’s point on the Dwarkesh podcast: export controls on Nvidia GPUs won’t stop China. They’ll just accelerate the development of AI that runs on Chinese chips.
new 1.6T total, 48B active open model coming very soon from @Meituan_LongCat
custom sparse attention, n-gram, adaptive expert computation, 1M context, 35T of training tokens, muon optimizer, multi teacher OPD, trained on 50k chinese gpus
https://longcat.chat/blog/longcat-2.0/
the new sparse attention method introduced with this model is basically a combination of components from existing ones. let's go over each sparse attention method and what they keep from them:
- deepseek sparse attention (DSA): they keep the top-k indexer, this is the basis of all the following sparse attention methods - DSA + index sharing (glm5.2): they share the index over multiple layers - minimax sparse attention (and also NSA from deepseek): they do block level top-k indexing -> they add another top-k at the token level to select precisely each token in the selected block - compressed sparse attention (from deepseek V4) and also NSA: they keep the sliding window -> they also add a sink token iirc
tl;dr: they do block level indexing, then token level on the selected blocks + add a sliding window and sink component with a 50/50 budget split. they share the token level top-k across layers
Introducing LongCat-2.0 🐱 1.6T parameters · MoE with ~48B active · 1M context The full model behind Owl Alpha on @OpenRouter — now available.
Built for agentic coding from the ground up: ◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens ◆ Zero-Compute Experts — dynamic activation 33B–56B per token, zero wasted compute ◆ MOPD — three specialized expert groups (Agent / Reasoning / Interaction), gate-routed per task
How it stacks up: → Terminal-Bench 2.1: 70.8 → SWE-bench Pro: 59.5 (GPT-5.5: 58.6) → SWE-bench Multilingual: 77.3 → FORTE: 73.2 · RWSearch: 78.8 · BrowseComp: 79.9
📖 Tech Blog: https://longcat.chat/blog/longcat-2.0/ Try it across different scenarios 🧵👇
le chaton long > le chaton fat for now
Introducing LongCat-2.0 🐱 1.6T parameters · MoE with ~48B active · 1M context The full model behind Owl Alpha on @OpenRouter — now available.
Built for agentic coding from the ground up: ◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens ◆ Zero-Compute Experts — dynamic activation 33B–56B per token, zero wasted compute ◆ MOPD — three specialized expert groups (Agent / Reasoning / Interaction), gate-routed per task
How it stacks up: → Terminal-Bench 2.1: 70.8 → SWE-bench Pro: 59.5 (GPT-5.5: 58.6) → SWE-bench Multilingual: 77.3 → FORTE: 73.2 · RWSearch: 78.8 · BrowseComp: 79.9
📖 Tech Blog: https://longcat.chat/blog/longcat-2.0/ Try it across different scenarios 🧵👇
🎁 Launch offer — limited time: ✦ 50M tokens for just $1.9 ✦ 1B tokens for $59.9 (80% off) Cache hits are free. Unused tokens fully refundable.
→ Try it now: https://longcat.chat/platform/product → Join our Discord for more demos, support, and direct feedback：https://discord.gg/jmuv5TQzKU
Damn, that's right. So: "≈48B" (thanks to N-Gram embedding, variable) active, 35T tokens. V4-tier, ≈8e24? Would be the biggest Chinese pretraining on domestic hardware. Some strange "Superpods": > "our accelerators" > "up to 48 machines each" > below 80 GB HBM per accelerator > "The device offers limited HBM bandwidth but a relatively large L2 cache" > "built-in 200 Gbps network adapter within the accelerator"
what is this stuff Was the first truly big Chinese model trained on domestic compute also using some extremely obscure piece of hardware?
@teortaxesTex @zugzwangg_ it said 35T+ tokens
💻 Coding: Reads your entire codebase at once — understands the architecture, tracks cross-file dependencies, and delivers changes that compile clean.
I was just mentioning today how counterintuitive it is because it pushes Chinese labs to be super creative around architecting models
Meituan, basically China’s DoorDash, trained a 1.6T parameter LLM on 50K Chinese chips.
It reminds me of Jensen Huang’s point on the Dwarkesh podcast: export controls on Nvidia GPUs won’t stop China. They’ll just accelerate the development of AI that runs on Chinese chips.
Hm, yes, it does match if you remember that 910C is two "910B" dies glued together. Okay then! 384 CloudMatrix – or rather, a hundred or so of them – is good enough to train a 1.6T MoE, it seems. This is 2025 tech. In ≈3 weeks we'll see actual "super"PODs.
@teortaxesTex These are almost definitely Huawei Ascend 910C. 910C superpod has 48 machines with 8 processors each. Each processor has two physical dies that could function as two logical processors. In that mode, each die has 64GB HBM and 200 Gbps RDMA networking so everything matches.
Clearer mark of a new era: latest models (here longcat) are only evaluated on agentic capabilities.
LongCat-2.0 dropping on Hugging Face soon
"Chinese DoorDash" releases a 1.6T model trained on 35T tokens on Chinese ASICs, reaching close to frontier performance.
Introducing LongCat-2.0 🐱 1.6T parameters · MoE with ~48B active · 1M context The full model behind Owl Alpha on @OpenRouter — now available.
Built for agentic coding from the ground up: ◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens ◆ Zero-Compute Experts — dynamic activation 33B–56B per token, zero wasted compute ◆ MOPD — three specialized expert groups (Agent / Reasoning / Interaction), gate-routed per task
How it stacks up: → Terminal-Bench 2.1: 70.8 → SWE-bench Pro: 59.5 (GPT-5.5: 58.6) → SWE-bench Multilingual: 77.3 → FORTE: 73.2 · RWSearch: 78.8 · BrowseComp: 79.9
📖 Tech Blog: https://longcat.chat/blog/longcat-2.0/ Try it across different scenarios 🧵👇
@Meituan_LongCat @OpenRouter IT'S HIM
📊 Analysis: Understands your question in natural language, plans the analysis steps, and delivers business insights — no SQL, no manual wrangling.
✍️ Writing: Handles long-form content with consistent structure, tone, and detail throughout, even across rewrites.
For the latest pricing and offer details, please refer to our official website.
📑 Research: Give it a topic. It searches, reads, cross-references, and refines autonomously until it delivers a structured review.
Introducing LongCat-2.0 🐱 1.6T parameters · MoE with ~48B active · 1M context The full model behind Owl Alpha on @OpenRouter — now available.
Built for agentic coding from the ground up: ◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens ◆ Zero-Compute Experts — dynamic activation 33B–56B per token, zero wasted compute ◆ MOPD — three specialized expert groups (Agent / Reasoning / Interaction), gate-routed per task
How it stacks up: → Terminal-Bench 2.1: 70.8 → SWE-bench Pro: 59.5 (GPT-5.5: 58.6) → SWE-bench Multilingual: 77.3 → FORTE: 73.2 · RWSearch: 78.8 · BrowseComp: 79.9
📖 Tech Blog: https://longcat.chat/blog/longcat-2.0/ Try it across different scenarios 🧵👇
