The editorial argues that Meituan's LongCat 2.0 release — a ~560B MoE with ~27B active parameters — makes the food-delivery company the second Chinese consumer-internet giant after ByteDance to ship frontier-scale open weights. The thesis is that companies with brutal unit economics in physical logistics can absorb $30M training runs as marketing spend, quietly shifting the open-weight leaderboard away from AI-native labs.
By publishing LongCat 2.0's weights on GitHub — following the 1.0 paper's benchmark claims against DeepSeek-V3 and Qwen2.5-Max — the Meituan team is signaling that a non-AI-native super-app can compete at frontier scale. The release itself is the position: open weights from a food-delivery company treated as normal infrastructure output.
The FuckClaude reverse-proxy exists because regional availability and payment friction — not capability — are what block developers from using Claude. Publishing a serious Python/Node middleware under an abrasive name and watching it trend to 366 stars asserts that the demand is for plumbing around geographic gatekeeping, not for a better model.
The chatgpt-register-sub2api repo automates registration and turns subscription accounts into API endpoints, echoing the same underlying claim: developers will build elaborate workarounds to route around US provider access gates. Its trending position (186 stars, 89 comments) validates that this is a widespread pain point, not an edge case.
The Codex-5.5-codex-instruct-5.5 repo topped the trending list at 1,360 stars by piggybacking on OpenAI's Codex brand and a plausible-looking version number. Its position at the top of the board — above a real frontier model release — is itself the argument that discovery on GitHub has become a naming-SEO game.
The editorial explicitly frames the three co-trending repos as 'three different signals in one screenshot,' with the name-squat repo out-starring the legitimate Meituan release by 6x. This treats GitHub's trending algorithm as a lagging, easily-manipulated indicator rather than a signal of actual technical importance.
Three repos hit GitHub trending in the same window, and read together they sketch the current shape of the open-weight LLM race better than any launch post.
The headline release: meituan-longcat/LongCat-2.0. Meituan — the Chinese super-app better known for delivering hotpot than shipping transformers — has open-weighted the successor to LongCat 1.0, a ~560B-parameter Mixture-of-Experts model with ~27B active parameters that landed in late 2025. The v2.0 repo is trending at 222 stars in its first push window, which understates it: MoE releases from non-AI-native companies typically get discovered on a delay, after the first wave of independent evals lands. Meituan is now the second Chinese consumer-internet giant, after ByteDance's Doubao team, to publish frontier-scale open weights — and neither of them lists 'foundation models' in their revenue mix.
Directly above it on the same trending list: LinXiaoTao/FuckClaude at 366 stars — a reverse-proxy that wraps Claude's API to sidestep regional availability and payment friction. And leading the board at 1,360 stars: yynxxxxx/Codex-5.5-codex-instruct-5.5, a name-squat repo trading on OpenAI's Codex brand. Three repos, three different signals, all in one screenshot.
The LongCat 2.0 release matters on its own terms. Meituan's 1.0 paper claimed competitive scores on Chinese-language reasoning benchmarks against DeepSeek-V3 and Qwen2.5-Max, at a training cost the team pegged in the low eight figures — cheap enough that a food-delivery company's ML team can justify it as infrastructure spend. The unspoken thesis of every Meituan-LongCat release is that if your unit economics survive delivering a $4 lunch across Beijing rush-hour traffic, you can absorb a $30M training run and treat it as a marketing expense.
But the more interesting signal is the co-trending. FuckClaude — abrasive name aside — is a serious piece of software: it's a Python/Node middleware that translates Anthropic's API surface into an OpenAI-compatible endpoint, letting Chinese developers plug Claude into stacks that already speak the OpenAI schema, often via routing through third-party billing intermediaries. Similar projects have existed for GPT-4 and Gemini for years. The star velocity on this one — 366 in a trending window — tracks with the recent Reuters report that Alibaba is internally restricting Claude Code over 'backdoor' concerns. When a state-adjacent enterprise blocks a Western model on security grounds, downstream devs don't stop using it — they just move to reverse-proxies, and the proxy authors trend on GitHub.
And then there's the Codex-5.5 repo. It's almost certainly not what it claims to be — OpenAI has not shipped a 'Codex 5.5,' and the repo owner's handle is a burner-style string of x's. But it's leading the trending list at 1,360 stars, which tells you exactly how much of GitHub trending in 2026 is legitimate signal versus SEO/name-squat/star-farm activity. GitHub trending is now a lossy channel — real frontier releases surface on the same page as astroturfed name-squats, and the ranking algorithm can't tell them apart.
Stack these together and you get the actual state of the ecosystem: (1) Chinese non-AI-native companies are joining the open-weights race with frontier-scale MoEs, (2) Chinese devs are aggressively routing around access restrictions to Western models via community-built proxies, and (3) the discovery layer where all of this is supposed to be visible — GitHub trending — is being flooded with noise from actors who've figured out it's cheaper to farm stars than to ship code.
If you're maintaining an AI-adjacent stack in mid-2026, three practical takeaways.
Re-run your open-weight evals every quarter, not every year. The LongCat 1.0 → 2.0 gap is measured in months, not release cycles. Qwen, DeepSeek, Kimi, Doubao, and now LongCat are all iterating faster than most in-house eval harnesses get re-run. If your last 'open weights aren't good enough' decision is more than six months old, it's a stale decision. Pin a recurring calendar item to re-benchmark on your actual workload, not on MMLU.
Treat community proxies as production risk, not a shortcut. FuckClaude and its cousins are convenient, but they insert an untrusted middleware layer between your keys and Anthropic. If any of your devs are running personal projects through a community proxy while logged into a corp account, you have a credential-exfiltration vector that no SOC2 audit will catch. Write it into your acceptable-use policy before it writes itself into your postmortem.
Stop treating GitHub trending as a signal. Use `github-trending-repositories` filtered by 'created in last 30 days' + 'stars in last 24h' + a manual sanity check on the README — trending on its own is now closer to Product Hunt than to a curated feed. The high-quality Chinese releases are increasingly announced on WeChat, Xiaohongshu, and Zhihu first, hit HuggingFace second, and only surface on GitHub trending third — by which point they're competing for attention with three name-squat repos and a Claude proxy.
The pattern to watch over the next two quarters is not 'when will an open model beat Claude on SWE-Bench' — that framing is already backwards. It's how many more non-AI-native Chinese companies (Meituan, ByteDance, and now speculatively JD, Pinduoduo, Xiaomi) treat foundation models as infrastructure they publish rather than SaaS they consume. Every one that ships is another data point that the moat around Western closed-model APIs is narrower than the pricing suggests. And every FuckClaude-style proxy that trends is another reminder that in a bifurcated internet, developers route around geopolitics faster than either side's policy teams can keep up.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.