Meituan LongCat 2.0 is HERE (Real Tests and Review)
Impressive recovery did not erase the failure
- LongCat 2.0 arrived with reported specifications of 1.6 trillion total parameters, 48 billion active parameters per token, and a one-million-token context window. Those numbers explain the interest; they do not prove coding reliability. (source video K_4kkcOzyKM, 00:35; 02:01)
- Ron passed its 3D ancient Chinese building despite an imperfect roof and non-working x-ray mode, but the model failed to make its simple space shooter playable on the first attempt. (source video K_4kkcOzyKM, 04:11; 04:34; 04:48; 05:04)
- The repair was better: LongCat identified a draw-function type error, fixed the dead animation loop in about two minutes, and produced a smooth playable game. The recovery did not restore Ron’s confidence in its one-shot ability. (source video K_4kkcOzyKM, 06:09; 06:18; 06:37; 07:20)
- Access was a serious practical constraint when filmed: Ron found cloud purchase limited to Alipay or WeChat Pay with a Chinese phone number, while the attractive token pack did not supply the API key needed for Claude Code. (source video K_4kkcOzyKM, 01:22; 01:33; 03:04; 03:13)
LongCat 2.0 is technically interesting, but it is not Ron’s daily driver and not an auxiliary agent he would add to his coding fleet on this evidence. (source video K_4kkcOzyKM, 07:43) The building demo earned a pass and the model diagnosed and repaired its broken shooter quickly, yet failing such a basic one-shot task left Ron’s confidence low. (source video K_4kkcOzyKM, 04:48; 06:09; 06:26; 06:37; 07:20) Treat this as a promising first test, not a complete capability map: Ron explicitly planned more testing and suspected its strengths might sit in codebase and migration-document work rather than creative game generation. (source video K_4kkcOzyKM, 08:38; 08:45)
Watch the test
- 00:00 · Why LongCat 2.0 matters: Ron introduces Meituan’s model and its reported scale.
- 00:55 · Specs and access limits: the open-source claim, the blocked download, cloud payment restrictions, and benchmark positioning.
- 02:06 · Running LongCat through Claude Code: Ron explains the Anthropic-compatible endpoint route and why the API key matters.
- 03:17 · Sparse attention and the context claim: the architecture claim behind the one-million-token window.
- 03:52 · 3D Chinese building test: clean motion and UI, but weak roof detail and incomplete feature behavior.
- 04:52 · Space shooter failure: a polished start button leads to a blank game.
- 05:57 · Second prompt and root-cause repair: the second prompt finds the type error and restores the animation loop.
- 07:26 · Coding verdict: why a successful repair still did not make LongCat fleet-ready.
- 09:02 · Recorded cost and final access warning: sub-cent requests shown in the test and the cloud-payment constraint.
Ron in his own words
“I would give it a pass for this project.” — Ron, source video K_4kkcOzyKM, 04:48
“And this is a big no no.” — Ron, source video K_4kkcOzyKM, 05:07
“at least it redeemed itself and uh fixed its problems, identified the root cause” — Ron, source video K_4kkcOzyKM, 07:00
“my confidence level is still not high.” — Ron, source video K_4kkcOzyKM, 07:25
A pass, a failure, and a recovery
The model card story and the operator story are different. The video describes sparse attention as LongCat’s alternative to standard dense attention. It says LongCat’s version reduces long-sequence computation from quadratic to linear complexity and enables a native one-million-token context window. That is the architecture explanation presented in the video, not an independently reproduced context test. (source video K_4kkcOzyKM, 03:17; 03:29; 03:33)
| Question | Evidence recorded in the video | Companion reading |
|---|---|---|
| Can it produce a polished visual demo? | The ancient-building animation and UI were smooth, but the roof had exposed scaffolding, x-ray mode failed, and weathering changed colour without adding texture or detail. (source video K_4kkcOzyKM, 04:06; 04:11; 04:34; 04:39) | Mixed pass. The main scene existed, but several hard details did not land. |
| Can it one-shot a simple game? | The shooter opened with a polished initialise button, then displayed no playable game. Ron calls that unacceptable for the task. (source video K_4kkcOzyKM, 04:55; 05:04; 05:07; 05:13) | No. Surface polish hid a blocking logic failure. |
| Can it debug its own output? | On the second prompt, it traced the failure to a draw function throwing a type error and killing the animation frame loop, then repaired the game in about two minutes. (source video K_4kkcOzyKM, 05:57; 06:09; 06:18; 06:26; 06:37) | Yes in this case. The repair loop was materially stronger than the first pass. |
| Was the recovered result usable? | The repaired game ran smoothly, with explosions, consistent texture, and movement drift that Ron liked. (source video K_4kkcOzyKM, 06:37; 06:44; 06:50; 07:12) | Playable after intervention, not one-shot reliable. |
The blank shooter was not presented as a VPS, browser, or payment failure. The reasoning trace identified a bug in the generated draw loop, and the same environment ran the repaired version. The separate cloud-payment and model-download restrictions affected access, not the quality of that generated game. (source video K_4kkcOzyKM, 01:04; 01:22; 06:09; 06:37)
Access and cost: keep the date attached
Ron used LongCat through Claude Code by pointing the harness at LongCat’s API base URL. He says the setup followed the same general method used with MiniMax and GLM, but the transcript does not preserve the exact endpoint, settings block, or launcher commands. This companion therefore does not invent a copy-paste configuration. (source video K_4kkcOzyKM, 02:06; 02:18; 02:53)
At filming, the token pack shown offered 50 million tokens for $1.90, but Ron says it did not provide an API key; Claude Code required pay-as-you-go API access instead. He also shows recorded requests costing less than one cent and describes the average as a fraction of a cent. These are screen-level observations from this run, not a current price sheet or a guarantee for a different workload. (source video K_4kkcOzyKM, 03:08; 03:13; 09:12; 09:17)
Before you use LongCat for real work
- Need dependable one-shot coding? Skip LongCat on this evidence. The first shooter output was not playable. (source video K_4kkcOzyKM, 05:04; 07:20)
- Comfortable with a repair loop? A limited trial is more defensible: the model found and fixed its generated-code error quickly. (source video K_4kkcOzyKM, 06:09; 06:26; 06:37)
- Evaluating long-context or repository work? Run a separate test. Ron proposed codebases and migration documents as a likely strength, but did not demonstrate that workload here. (source video K_4kkcOzyKM, 08:38)
- Outside the supported payment route? Verify access before comparing capability. At filming, cloud purchase required Alipay or WeChat Pay plus a Chinese phone number. (source video K_4kkcOzyKM, 01:22; 01:33)
Freshness note
The video was published June 30, 2026, and this companion was checked on July 18, 2026. The access restrictions, API-versus-token-pack distinction, and prices above are filming-time observations. Current repository availability, API access, payment methods, pricing, and later LongCat results were not independently refreshed, so verify those before acting on them. (source video K_4kkcOzyKM, 01:04; 01:22; 03:04; 03:13)