Longcat-2.0 is interesting. 1.6T model, roughly frontier quality though not near the top, but all trained on custom ASICs. Attempting to break the tyranny of training compute shortages.
6:01 AM · Jun 30, 2026 · 52 Views
Agility Robotics AI lead Chris Paxton details Longcat-2.0, a 1.6-trillion-parameter model trained entirely on custom ASICs
Story Overview
The headline frames Agility Robotics AI lead Chris Paxton as the source detailing a 1.6-trillion-parameter model trained solely on custom ASICs, yet available evidence attributes LongCat-2.0 to Meituan's team with no documented connection to Paxton or the robotics firm, leaving the reported origin unsupported.
0310131
Open Question
Who actually shipped the weights?
Meituan open-sourced the full 1.6T MoE model with 1M context, LongCat Sparse Attention, and N-gram embeddings after pretraining on over 50K domestic ASICs, releasing weights on Hugging Face under MIT license.
Pricing Watch
Benchmarks land near the frontier at flash-sale rates
LongCat-2.0 posts 59.5 on SWE-bench Pro and strong agentic scores while offering pay-as-you-go API pricing at $0.30 per million input tokens during the launch promo, with context caching free.
Sentiment
Users are optimistic about Longcat-2.0's frontier-scale training run on custom ASICs because it could ease compute shortages and help new entrants challenge established labs.
Pos
100.0%
Neg
0.0%
1 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS77LIKES2RETWEETS1
Chris Paxton@chris_j_paxton
A frontier-scale, nearly frontier quality, training run that is entirely on custom ASICs would make it much easier for new entrants to challenge the big players, especially in an environment like coding where I think building a good harness + RL training can make a huge difference vs pure data scaling
Chris Paxton@chris_j_paxton
Longcat-2.0 is interesting. 1.6T model, roughly frontier quality though not near the top, but all trained on custom ASICs. Attempting to break the tyranny of training compute shortages.
1hViews 77Likes 2Bookmarks 0
Noah Vandal@noah_vandal
@chris_j_paxton Interesting to see if this trend prevails
1hViews 13Likes 1
sophia@cis_female
@chris_j_paxton i find this confusing. anthropic/google have been using TPUs for years? anthropic has been using trainiums at least for inference as well.
Chris Paxton@chris_j_paxton
Longcat-2.0 is interesting. 1.6T model, roughly frontier quality though not near the top, but all trained on custom ASICs. Attempting to break the tyranny of training compute shortages.
1hViews 2Likes 0Bookmarks 0