AI Judge changed title after evaluation, original title: "Meituan unveils LongCat-2.0, an upcoming 1.6-trillion parameter open-source MoE model with a 1-million token context window"
Meituan's LongCat team just dropped an open-weight MoE model that scales to 1.6 trillion total parameters while activating roughly 48 billion per token, paired with a full million-token context window and training runs that stretched across more than 50,000 domestic Chinese AI ASICs in superpod setups.
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/
The entire pretraining run used over 35 trillion tokens on those ASIC clusters with custom 6D parallelism and no reported major rollbacks, though the precise ASIC model and superpod configuration details stay light in the announcement.
LongCat-2.0 adds sparse attention tricks, N-gram embeddings, zero-computation experts, and post-training specialist groups aimed at coding and agent workflows, with charts comparing it to Gemini 3.1 Pro, GPT-5.5, and Opus variants on Terminal-Bench, SWE-bench, and similar suites.
Some users praise Meituan's LongCat-2.0 training achievements and tech report on domestic ASICs while others criticize the model's image quality, Qwen integration, and inference readiness as subpar.
No Digg Deeper questions have been answered for this story yet.
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
https://longcat.chat/blog/longcat-2.0/
People are missing out on how big a deal Longcat 2.0 by Meituan (aka "Chinese Doordash") is.
Near frontier performance, trained on 50k Chinese domestic accelerators! The first ever to achieve this!
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/
910B had 64Gb already, and I don't think there exist 910B "superpods". @zephyr_z9 @tphuang I missed these news, they're not even very new. anyway, very cool development on N-grams
tbh as usual very interesting technical details from Meituan but when I click "try it" I get something like Qwen-4B, seemingly worse than their first LongCat-Flash. Completely fails in Russian too. Idk if that's their new model
tbh as usual very interesting technical details from Meituan but when I click "try it" I get something like Qwen-4B, seemingly worse than their first LongCat-Flash. Completely fails in Russian too. Idk if that's their new model
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?!
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.
@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.
DS and Enthropic suspect Enflame 邃思 2.0
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?
@zephyr_z9 @tphuang it's strangely identical to V4 in dimensions, and they boast of total determinism too. Replicating V4 in 2 months is a bit rich… but that's not even what happened if LongCat 2.0 is Owl Alpha – it was debuted on OR on Apr 28. What is going on here.
910B had 64Gb already, and I don't think there exist 910B "superpods". @zephyr_z9 @tphuang I missed these news, they're not even very new. anyway, very cool development on N-grams
> For numerical reliability, we rework a range of foundational operators to improve precision — for example, all reduction-type operators adopt a binary-tree segmented accumulation strategy to reduce floating-point error accumulation. We further validate the accelerator's compute precision under real-world LLM workloads against a strict high-precision baseline, confirming its arithmetic integrity and production readiness, and introduce bit-flip detection in selected compute-intensive operators to promptly catch anomalies caused by hardware bit flips.
Sounds like alien techs
@teortaxesTex okay so I did a little check on them. It seems like they have invested into MetaX & Moore Threads. So those are some possibilities.
@teortaxesTex > So: "≈48B" (thanks to N-Gram embedding, variable) a this is because of zero computation expert imo, not ngram (from previous tech report)
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 Is this the Engram from Deepseek's paper?
That's a Huawei Ascend 910C in a CloudMatrix 384.
(48 nodes X 8 acccelerators; calling it "supernode" and "superpod"; CANN's "graph mode"; < 80GB (910C has ≈64GB); etc)
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?
@yzhang_cs if only the model was half as good as yours
@Nacho_rodpe No, different but similar inspiration they already have published a model using it
@zugzwangg_ @teortaxesTex Clearly 384s Btw, rumors long cat 2.0 being trained on Ascend has been floating around for over a month U can find it on twitter too
@teortaxesTex not sure if relevant but i recall this from last RL paper
@eliebakouch @Meituan_LongCat I LOVE LONGCAT MODELS
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/
@andrew_n_carr @Meituan_LongCat You know what I love more than longcat models? Longcat tech report 😍😍😍
@eliebakouch @Meituan_LongCat I LOVE LONGCAT MODELS
