NEW 1.6 TRILLION Open Source AI vs Claude & GLM | LongCat-2.0 🫣
Credibility score: 50/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "NEW 1.6 TRILLION Open Source AI vs Claude & GLM | LongCat-2.0 🫣" and rated it 50/100 for credibility (a BS score of 50/100 — mixed credibility), on 2026-07-06. Its weakest claim — "Benchmarking against specific models but omitting others 🚩" — scored 20/100 and was flagged as cherry-picked. 21 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 21 claims analyzed: 1 scored under 40, 16 between 40 and 69, and 4 at 70 or above.
Claims analyzed
Video opens with a highlight reel preview, setting up a big claim about model performance. — Just Vibes (50/100)
At 0:00
Starting with a bold claim about 'no model will get this' is pure hype-building for the intro. Get ready for the main event! 🚀
Why this score: The opening line is a classic teaser, designed to grab attention and build anticipation for the 'Long Cat 2.0' model. It's not a factual claim to be verified, but a rhetorical device to engage the viewer.
Original quote: “And this is probably a test that I don't think any model will get at this. Hey guys, welcome to the show. Today we're checking out the Deliveroo of Sheena. We're checking out Long Cat 2.0.”
Claims Long Cat 2.0 is a 'massive openweight model' with '1.6 trillion total parameters'. — No Frame (75/100)
At 0:09
A 1.6 trillion parameter model is genuinely massive in the AI world. The number itself is verifiable. 🤯
Why this score: The speaker states the model's size in parameters, which is a specific, quantifiable detail. While 'massive' is subjective, the parameter count provides a concrete measure. This is a straightforward statement of fact about the model's architecture.
Original quote: “This is a massive openweight model. It is 1.6 trillion total parameters. They saw what Deep Seek has done and they figured, can we beat them?”
Mentions the full weight version was pulled, only 8-bit quantization released, with uncertainty on why. — Missing Context (45/100)
At 0:19
They dropped a huge model, then pulled the full version and just shrugged. 'I don't know' isn't exactly confidence-inspiring for a trillion-parameter release. 🤷♀️
Why this score: The speaker acknowledges a significant piece of missing information: why the full-weight model was pulled and only an 8-bit quantized version was released. This lack of explanation leaves a gap in understanding the project's status and the developers' intentions, which could impact perceptions of the model's stability or future. It's a direct admission of missing context.
Original quote: “And originally they were meant to be releasing the full weight version of this, but they took down the repo. I don't know if they're still uploading or what they decided, but what they have released is the 8bit quantization version of it, but”
Explaining memory requirements for running the model locally 💾 — No Frame (75/100)
At 0:31
Just laying out the specs for running this beast locally. Two terabytes for 8-bit, one for 4-bit. That's a lot of RAM, chief. 🤯
Why this score: The speaker is providing technical specifications for running the AI model, specifically memory requirements for different quantization levels. This is straightforward information, not a rhetorical trick.
Original quote: “nonetheless, 8 bit quantization is 2 terb of just storage. So, you need to have over two terabytes of memorandios if you're going to run it at 8 bit. And if you if you try running at 4bit, which everyone seems to love that bit quantization, that's the terabyte of memory.”
Throws out massive storage requirements with casual confidence. 🤯 — Confidence Mismatch (45/100)
At 0:31
Two terabytes for 8-bit, one for 4-bit, just like it's a USB stick. The numbers are huge but the delivery is chill. 💾
Why this score: The speaker states very large memory requirements (2TB for 8-bit, 1TB for 4-bit) for running the model locally. While these numbers might be accurate for the model, the casual way they're presented, as if it's a minor detail, creates a confidence mismatch. For most users, 1-2 terabytes of RAM is an astronomical amount, not a 'just' storage figure.
Original quote: “nonetheless, 8 bit quantization is 2 terb of just storage. So, you need to have over two terabytes of memorandios if you're going to run it at 8 bit. And if you if you try running at 4bit, which everyone seems to love that bit quantization, that's the terabyte of memory. So,”
Benchmarking against specific models but omitting others 🚩 — Cherry-Picked (20/100)
At 1:41
They're showing benchmarks against some models, but conveniently skipping the 'current top two' in his opinion. Selective data much? 🍒
Why this score: The speaker points out that while the model is benchmarked against Gemini 3.1 Pro and various Opus versions, it's not tested against what the speaker considers the 'current top two models' (Kimik K 2.7s and GLM5.2). This omission could present a more favorable comparison by excluding stronger competitors.
Original quote: “So, in their benchmarks over here, they're testing it against Gemini 3.1 Pro, Jipped, Opus, the old one, and the new one. Oh, they got all the opuses out. So, it's better than Gemini definitely. You can see that even in SWE, but it's just above Opus 4.6, just below 4.7. They haven't actually tested…”
Declares it 'definitely better' than Gemini, then immediately qualifies it. 🚩 — Volume Game (45/100)
At 1:50
Starts with 'definitely better' than Gemini, then immediately says it's 'just above' Opus 4.6 and not even tested against the top two. That's a quick walk-back. 🚶♂️
Why this score: The speaker makes a strong, definitive claim ('better than Gemini definitely') but then immediately softens it by saying it's 'just above' another model and hasn't even been tested against what they consider the 'current top two models'. This is a classic volume game: make a bold claim, then quietly retract or qualify it in the same breath, hoping the initial punchy statement sticks.
Original quote: “So, it's better than Gemini definitely. You can see that even in SWE, but it's just above Opus 4.6, just below 4.7. They haven't actually tested out against the current top two models in my opinion.”
Praising a Chinese delivery company's AI models despite their origin 🇨🇳 — No Frame (75/100)
At 2:02
Giving props to a Chinese delivery company for their 'impressive work' in AI models. Good tech is good tech, no matter where it's from. 👏
Why this score: The speaker acknowledges the origin of the models ('delivery of China') but focuses on the quality of their work ('impressive work', 'good models'). This is a straightforward assessment of the technology, not a manipulative framing technique.
Original quote: “Anyway, the Kimik K 2.7s and the GLM5.2 not there, but they're talking about the architecture. And if you have checked out long catat's models, I know they're the delivery of China, but they have made some impressive work. You check out the long cat flash, all these kind of good models they've done…”
Acknowledges 'delivery of China' then pivots to 'impressive work.' 🇨🇳 — Missing Context (45/100)
At 2:08
Mentions 'delivery of China' like it's a footnote, then immediately praises their 'impressive work.' Skips over any potential implications. 👀
Why this score: The speaker briefly mentions 'they're the delivery of China' in reference to LongCat's models, then quickly moves to praise their 'impressive work.' This phrasing could be a subtle way to acknowledge the origin without delving into any potential geopolitical or data privacy concerns that might be associated with Chinese-developed AI models, thus missing context that some viewers might find relevant.
Original quote: “And if you have checked out long catat's models, I know they're the delivery of China, but they have made some impressive work. You check out the long cat flash, all these kind of good models they've done in the past.”
Setting up a comprehensive comparison for model competence 🧪 — No Frame (75/100)
At 2:25
He's gonna test this model against the 'best local generation' and cloud versions. This is how you actually test things, folks. Let's see it! 💪
Why this score: The speaker is outlining a clear and comprehensive testing methodology, comparing the new model against established 'best local generation' models (GLM 5.2, Kimik 2.7) and cloud versions. This indicates a genuine attempt to assess competence, not a framing trick.
Original quote: “So, the first prompt I'm going to ask you is, is 3D platform game like Super Mario, and we're going to be testing its results with the best local generation. I'm talking about GLM 5.2, Kimik 2.7, also going to be putting it up against the cludes and the cloud versions of the model to see if it…”
Comparing LongCat-2.0's game generation to Gemini and Opus, calling Gemini's 'horrible' and Opus's potentially worse with thinking disabled. — Confidence Mismatch (45/100)
At 2:30
Calling Gemini's generation 'horrible' and Opus's potentially worse, but it's all based on a single, subjective test run. That's a strong take for one game. 🎮
Why this score: The speaker is making a strong, subjective judgment ('horrible,' 'better') based on a single instance of game generation. While they can have an opinion, presenting it as a definitive comparison without more rigorous testing or objective metrics is a confidence mismatch. It's just one game, dude.
Original quote: “It's definitely a better gen generation than Gemini. Gemini made a horrible generation. It might even be better than what Opus made when we had thinking disabled.”
Compares LongCat-2.0's generation to Gemini and Opus, calling Gemini's 'horrible' and LongCat's 'better'. — Loaded Language (45/100)
At 2:30
Calling Gemini's output 'horrible' is a vibe, not a metric. It's just a subjective opinion dressed up as a definitive judgment 💀.
Why this score: The speaker uses strong, subjective language ('horrible') to describe Gemini's generation, which is an emotional appeal rather than an objective comparison. While they show the output, the judgment itself is loaded, not a neutral assessment.
Original quote: “It's definitely a better gen generation than Gemini. Gemini made a horrible generation. It might even be better than”
Dismissing 'LongCat' for coding despite claims it's 'better than Gemini' — a classic 'trust my gut, not the marketing' move 🤷♂️ — Confidence Mismatch (45/100)
At 4:30
The speaker's personal experience with 'LongCat' for coding contradicts its advertised superiority over Gemini. It's a 'show, don't tell' moment where the 'show' isn't impressive. 📉
Why this score: The speaker expresses a personal opinion ('maybe isn't the best for coding') that directly contrasts with an unverified claim about the AI model ('it says it's better than Gemini'). This highlights a confidence mismatch between the model's self-proclaimed capabilities and the user's practical experience, without providing specific evidence for either side beyond personal anecdote.
Original quote: “So, I'm thinking this long cat maybe isn't the best for cuding even though it says it's better than Gemini, but or maybe some different cases.”
Doubts LongCat's coding ability despite claims of being 'better than Gemini'. — Confidence Mismatch (45/100)
At 4:30
Dismissing 'better than Gemini' with a 'maybe isn't the best' after one bad example. That's a quick judgment call. 🚩
Why this score: The speaker is expressing skepticism about LongCat's coding performance based on a single, possibly flawed, test. They acknowledge the model's claim of being 'better than Gemini' but quickly dismiss it without comprehensive testing, showing a mismatch between their limited evidence and their conclusion.
Original quote: “So, I'm thinking this long cat maybe isn't the best for cuding even though it says it's better than Gemini, but or maybe some different cases.”
Declares an answer, then immediately calls it 'debatable.' The confidence whiplash is real. 🎢 — Volume Game (45/100)
At 6:30
Says it's 'the answer' then immediately 'debatable.' The conviction lasted like 2 seconds. 💀
Why this score: The speaker presents a definitive answer ('the answer is a thousand') and then immediately undermines it by stating it's 'debatable.' This rhetorical move is a classic volume game, where a bold claim is made, then quietly walked back, allowing the initial confidence to linger while providing an escape hatch.
Original quote: “Now, this is the debatable what the answer is, but we're going to see what it comes down to.”
Declares an answer, then immediately says it's debatable. The old volume game. 📢 — Volume Game (45/100)
At 6:30
Drops a definitive answer, then immediately hedges it. The confidence is a rental. 🤡
Why this score: The speaker states 'the answer is a thousand' with certainty, then immediately undermines it by saying 'this is debatable'. It's a classic volume game: make a bold claim, then quietly retract or qualify it, hoping the initial claim sticks.
Original quote: “Now, this is the debatable what the answer is, but we're going to see what it comes down to.”
Claiming models are 'very very expensive to run' without specific costs 💸 — Confidence Mismatch (45/100)
At 8:30
Says 'very very expensive' after blowing $5 on three prompts. That's not 'very very expensive,' that's a coffee. ☕
Why this score: The speaker uses strong, definitive language ('very very expensive') to describe the cost of running AI models, but the only evidence provided is spending $5 on a few prompts. This is a mismatch between the confidence of the claim and the actual data presented, which doesn't support the 'very very expensive' assertion for general use.
Original quote: “So these these models are very very expensive to run and he doesn't even get”
AI got lost and started talking about witches instead of pronunciation. Classic AI tangent 🧙♀️ — Just Vibes (50/100)
At 10:30
The AI went full 'Hansel and Gretel' when asked for pronunciation. Not a claim, just a funny fail 💀
Why this score: The speaker is just narrating the AI's amusing failure to perform the task, veering off into unrelated topics like 'witches and good and wicked.' It's a humorous observation, not a factual claim to be verified.
Original quote: “just it got a bit lost and now I don't know, it's telling me about witches and good and wicked. It's not really telling me how to pronounce which each other. So that's failed miserably.”
DeepSync took 30+ seconds to say the pronunciation is 'how you already say it,' then called it a joke. — Confidence Mismatch (45/100)
At 10:30
30 seconds to say 'you already know it,' then 'it's a joke'? That's a lot of processing for a non-answer. 🤡⏳
Why this score: The AI model (DeepSync) took a significant amount of time (32-34 seconds) to produce an answer that essentially deflects the question back to the user, then the speaker claims the AI thinks it's a joke. This shows a mismatch between the effort (thinking time) and the quality/certainty of the answer, implying the AI is confident in a non-answer or a misinterpretation.
Original quote: “And I thought for 32 34 seconds. It's a lot of thinking to come up with this. And it says it pronounces it. It's pronounced the way you've already pronounced it in your head. So the answer is simply how you just did it. And that is probably the right answer because that is how I pronounced that. I…”
Suggests AI improvement paths, making a speculative leap — Confidence Mismatch (45/100)
At 12:30
He's confidently suggesting specific fixes like 'verbosifizing the transcript' as if it's a known solution, not just a guess. 🔮✨
Why this score: The speaker is speculating on potential future improvements for AI models, suggesting methods like 'feeding it multimodal data' or 'verbosifizing the transcript.' While these are plausible ideas, they are presented with a degree of certainty ('that's something they can put in their training data and maybe then things will get GI') that outpaces the current evidence or established solutions. It's a confident projection rather than a proven path.
Original quote: “So, let's see if these models improve in the future, understanding how to annunciate the words like the annunciation is done or will they need to start feeding it multimodal data or actually maybe verbosifizing the the transcript because you're just getting subtitles but you don't know how to…”
Acknowledges current model flaws but pivots to future 'flash version' improvements. — Volume Game (45/100)
At 14:30
Admits the current version is 'not beating' others, then immediately hypes up a future 'flash version.' Classic 'it's bad now but wait!' move 🙄
Why this score: The speaker uses a volume game by quickly acknowledging the current model's shortcomings ('not beating GLM and Kimmy', 'a bit verbose') but then immediately shifts focus to a promised, improved 'flash version' to maintain enthusiasm. It's a common sales tactic to downplay present issues by hyping future solutions.
Original quote: “And maybe it's not beating GLM and Kimmy at the moment and maybe it is a bit verbose, but these guys, they always release a flash version. They haven't like given up on the smaller bottle versions.”
See the full analysis with timestamps →