NEW Open Source Qwen? BEATS GLM 5.2 & Claude? 🧐 Ornith 397B TESTED
Credibility score: 48/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "NEW Open Source Qwen? BEATS GLM 5.2 & Claude? 🧐 Ornith 397B TESTED" and rated it 48/100 for credibility (a BS score of 52/100 — mixed credibility), on 2026-06-28. Its weakest claim — "Claiming OR N 'promises to be the best state-of-the-art coding agent' based on 'amazing' benchmarks." — scored 45/100 and was flagged as confidence mismatch. 20 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 20 claims analyzed: 0 scored under 40, 18 between 40 and 69, and 2 at 70 or above.
Claims analyzed
Opening with a random mask bit and 'dudy stuffs' for rich people. Just vibes. — Just Vibes (50/100)
At 0:00
Starting off with a completely random, unhinged intro about masks and 'dudy stuffs' — pure chaos for engagement 🎭
Why this score: The speaker is using an absurd, attention-grabbing opening that has no direct relevance to the video's topic. It's a comedic, 'just vibes' moment designed to hook the viewer rather than make a substantive claim.
Original quote: “And what is this? Is this like a mask you wear here? Do some dudy stuffs that the rich people do. Hey guys, welcome to the show.”
Claiming OR N 'promises to be the best state-of-the-art coding agent' based on 'amazing' benchmarks. — Confidence Mismatch (45/100)
At 0:16
Calling it 'the best state-of-the-art' and 'amazing' right out the gate, before showing any actual data. That's a lot of hype for a 'promise' 🚀
Why this score: The speaker is using highly confident and superlative language ('best state-of-the-art', 'amazing benchmarks') to describe a new model without yet presenting the evidence. This creates a strong positive bias before the viewer can evaluate the claims independently. The 'promise' is being presented as a near-certainty.
Original quote: “Hey guys, welcome to the show. Today we're checking out OR N which is a fine tune on Quen 3.5 and it promises to be the best state-of-the-art coding agent out there and they have some benchmarks and the benchmarks look amazing. Check this out. They're”
Claiming 'shady' behavior because a benchmark was omitted, implying malice. — Loaded Language (45/100)
At 0:30
Omitting a benchmark is one thing, but jumping straight to 'shady' and 'hiding' is a leap. Maybe it was an oversight? 🕵️♂️
Why this score: The speaker uses emotionally charged words like 'shady' and 'hiding' to suggest malicious intent without presenting any direct evidence beyond the omission itself. This frames the omission as deliberate deception rather than a potential error or different testing methodology.
Original quote: “But I noticed that in terminal bench they have omitted GLM's performance. So GLM actually got 81. So there might be something shady going on because they're hiding those benchmarks.”
Points out GLM's benchmark omission, implying shadiness. — Missing Context (45/100)
At 0:30
Noticed a benchmark was conveniently left out — that's like showing your best angle and cropping out the double chin 😬
Why this score: The speaker highlights that GLM's performance was omitted from the terminal bench results, suggesting a deliberate act to hide less favorable comparisons. This is a classic missing context move, making the new model look better by removing a strong competitor from the comparison.
Original quote: “in SME Bench Pro, you can see that is scoring higher than GLM 5.2. And GLM 5.2 to his amazing model since being in it. But I noticed that in terminal bench they have omitted GLM's performance. So GLM actually got 81. So there might be something shady going on because they're hiding those benchmarks.”
Qwen's less tokenized result is better, but acknowledges luck with the seed. — Volume Game (45/100)
At 2:30
Declares Qwen's result 'better' then immediately hedges with 'could have just been lucky with the seed.' The confidence giveth, and the confidence taketh away 🎲
Why this score: The speaker makes a strong claim about Qwen's superior performance ('this result is better') despite using fewer tokens, but then immediately walks it back by suggesting it might be due to 'luck with the seed.' This is a classic volume game, where a bold assertion is made and then quietly undermined, allowing the positive impression to linger while providing an escape clause.
Original quote: “So even though it's a less tokenized generation, I can say that this result is better than the result the massive result that the onif one get. But again, we could have just been lucky with the seed.”
Ornith model got into a runaway loop with thinking disabled, so it 'didn't count' 🤡 — Missing Context (45/100)
At 2:30
Dismissing a failure because 'it didn't count' is a convenient way to ignore bad results. The loop happened, regardless of the 'rules' 🙄
Why this score: The speaker is downplaying a clear failure of the Ornith model by saying it 'didn't count' because it went into a runaway loop, even though 'thinking was disabled'. This conveniently removes a negative result from the comparison without fully explaining why it shouldn't be considered, or why 'thinking disabled' led to a loop.
Original quote: “Now I asked or and it even though I had thinking disabled, it went ahead and did a lot of thinking and it ended up getting into a runaway loop. So that didn't count.”
Giving specific token counts and speed for a model run, but the visual doesn't match the instruction. — Confidence Mismatch (45/100)
At 4:30
He's rattling off numbers like they're gospel, but the visual on screen is clearly not doing what he asked. The cat's still too close! 🤦♂️
Why this score: The speaker states precise token counts and processing speed, implying a successful execution of instructions. However, he immediately follows up by noting the camera hasn't moved back as instructed, creating a mismatch between the confident technical data and the observed outcome. The numbers are real, but the 'success' they imply is not fully realized.
Original quote: “And boom. 3,167 tokens with thinking enabled at 21 tokens a second. Let's see if this is my instructions.”
Suggests finetuning might have overfit to pass benchmarks. — Confidence Mismatch (45/100)
At 4:30
Went from 'maybe' to a pretty confident 'overfit somehow' without any actual evidence. That's a leap of faith, chief. 🤸♀️
Why this score: The speaker starts with 'maybe' but then presents a specific explanation ('overfit somehow to pass the benchmarks') as a likely cause without providing any data or analysis to support this particular hypothesis. It's a speculative jump presented with more certainty than warranted by the evidence shown.
Original quote: “So it couldn't handle the heat there. So maybe the finetuning the extra data that you feed might have overfit somehow to pass the benchmarks that they made.”
Randomly linking a mask to 'Trumpian parties' and 'dodgy stuffs' — Loaded Language (45/100)
At 6:30
Went from 'what is this mask?' to 'Trumpian parties' and 'dodgy stuffs' real quick. That's a leap, not a logical step 🤡
Why this score: The speaker sees a mask on screen and immediately jumps to associating it with 'Trumpian parties' and 'dodgy stuffs that the rich people do,' using emotionally charged language without any actual evidence or context for the mask's purpose. It's pure speculation dressed up with loaded terms.
Original quote: “what is this? Is this like a mask you wear? We doing some of those uh Trumpian”
Comparing models, saying smaller is better, but where's the proof? 🧐 — Confidence Mismatch (45/100)
At 8:30
Claims the smaller version is better without showing the 'big guy's' results for comparison. Just vibes, no data. 🤷♀️
Why this score: The speaker asserts the 3.6 smaller version produces better results than the 'big guy' without providing any visual or statistical evidence for the 'big guy's' performance. It's a confident statement lacking a direct comparison.
Original quote: “Even the big guy doesn't produce as good results as the 3.6 smaller version.”
Claiming the smaller 3.6 version beats the 'big guy' on Flappy Birds. — Confidence Mismatch (45/100)
At 8:30
Saying the smaller version is better, but then the demo shows it's 'a bit broken.' The confidence is doing overtime here 🤡
Why this score: The speaker confidently states the smaller 3.6 version produces better results than the 'big guy' (presumably a larger model), but immediately follows up by showing the 3.5 version (which is implied to be similar in performance to the 3.6 for this task) is 'a bit broken' and doesn't perform well. This creates a mismatch between the initial confident claim and the immediate visual evidence/admission.
Original quote: “Even the big guy doesn't produce as good results as the 3.6 smaller version.”
Claiming 'thinking enabled' is essential for general knowledge tasks. — Confidence Mismatch (45/100)
At 10:30
Declaring 'thinking enabled' is a MUST for general knowledge. Bro, it's a setting, not a magic wand. 🧙♂️⚙️
Why this score: The speaker states with high confidence that 'thinking enabled' is a requirement for general knowledge tasks. While it might improve performance, framing it as an absolute necessity without further evidence or nuance is a bit of a leap.
Original quote: “With thinking enabled, it does know. So you need to have if you want to come up with these general knowledge, you need to make sure you got the thinking stuff enabled, not for things.”
AI knows Fort Minor, not Fallout Boy, with 'thinking enabled' — a specific capability claim. — No Frame (75/100)
At 10:30
The AI correctly identified Fort Minor when 'thinking enabled' was on. That's a win for the AI, I guess. 🤖
Why this score: The speaker is demonstrating a specific capability of the AI model (Qwen) when a certain feature ('thinking enabled') is active. This is a direct observation of the AI's performance in a controlled test, not a rhetorical trick. It's a straightforward report of a test result.
Original quote: “Fallout Boy. No, it's not the Fallout Boy. But with thinking enabled, Quen knows that it's Fort Miner. Remember the name.”
Dismisses models as 'fail' without showing the actual 'fail' 💀 — Missing Context (45/100)
At 12:30
Calls it a 'fail' but we don't see the actual output that failed. Just trust me, bro energy. 🤷♂️
Why this score: The speaker labels a previous test as a 'fail' and compares it to other models without showing the specific output or criteria for failure. This leaves the viewer without the necessary context to understand the judgment.
Original quote: “fail. And it's nowhere near as good as
the Kimmy K2.7s and the the GLM's. And”
Calls a car animation 'definitely a good generation' — then immediately backtracks — Volume Game (45/100)
At 12:30
Starts with 'definitely a good generation' then immediately pivots to 'going a bit off road' and 'quality loss.' The confidence evaporated fast 💨
Why this score: The speaker initially makes a strong, positive claim about the animation's quality ('definitely a good generation') but then immediately qualifies it with significant drawbacks ('going a bit off road,' 'quality loss'). This is a volume game: a loud positive claim followed by a quiet retraction.
Original quote: “This one here is to do a canvas
animation and it's making a car on the
road. And that is definitely a good”
Bragging about 25 tokens/second speed for Super Mario, but the 'thinking disabled' part is sus 🚩 — Missing Context (45/100)
At 14:30
Yeah, it's fast when you turn off its brain. That's like saying my car is fast if I take out the engine and push it. 🏎️💨
Why this score: The speaker highlights the speed (25 tokens/second) but conveniently downplays that this performance is achieved with 'thinking disabled.' This omits crucial context about the model's actual capability when fully engaged, making the 'super fast' claim less impressive.
Original quote: “Super Mario. Let's see if it can make Super Mario thinking disabled. 6,000 tokens. Now, this is fast. 25 tokens second. Super fast.”
Bragging about speed with 'thinking disabled' — that's a choice 🏎️💨 — Missing Context (45/100)
At 14:30
Calling 25 tokens/second 'super fast' for a model with 'thinking disabled' is like praising a car for speed when it's on a flatbed. Context matters! 🤷♂️
Why this score: The speaker highlights the speed (25 tokens/second) of the model when 'thinking disabled,' framing it as 'super fast.' However, disabling the 'thinking' process would naturally make a model faster as it bypasses complex reasoning, which is a crucial piece of context missing from the 'super fast' claim. It's like comparing a stripped-down race car to a fully loaded one without mentioning the missing features.
Original quote: “Let's see if it can make Super Mario thinking disabled. 6,000 tokens. Now, this is fast. 25 tokens second. Super fast.”
Exaggerating token usage for a simple task — the 'Miniax went to the max' bit 💀 — Loaded Language (45/100)
At 16:30
100,000 tokens for a train schedule? That's a lot of words for a timetable, sounds like a bit of an overstatement for comedic effect. 😂
Why this score: The speaker uses hyperbole ('100,000 tokens regurgitating the Amtrak train schedule') to emphasize the inefficiency of the model, creating a humorous, relatable frustration rather than a precise technical claim. It's a 'Miniax went to the max' joke.
Original quote: “some of them even like Miniax M3 spent 100,000 tokens tokens. I can't even say that word. 100,000 tokens regurgitating the Amtrak train schedule and you didn't even get to writing code. Just just write I had to stop it. It was too much miniax. It went to the max.”
GLM 5.2 suggested 'cut the children' as a top token, Ornith doesn't. — No Frame (75/100)
At 18:30
Comparing model outputs on a specific prompt — the GLM 5.2 result is wild, and Ornith avoids it. Good catch! 💀
Why this score: The speaker is directly comparing the behavior of two different AI models (GLM 5.2 and Ornith 397B) when given a specific prompt about distributing oranges with a knife. He highlights a concerning output from GLM 5.2 ('cut the children') and notes that Ornith 397B does not exhibit this behavior. This is a straightforward observation of model performance, not a rhetorical trick.
Original quote: “So, give is still good. We saw with GLM 5.2 the third most popular token was to cut. And when you chose that token, it said cut the children. So yeah, and that was the third most topseeking token. It was only 3%. But it was the third on the list. This guy here, you look through give, there's no…”
Describes a 'gorgeous generation' of code, then notes a random output difference based on 'thinking enabled/disabled'. — Confidence Mismatch (45/100)
At 20:30
Calls it 'gorgeous generation' then immediately points out a weird, inconsistent output. The confidence in 'gorgeous' is doing some heavy lifting there 🤡.
Why this score: The speaker initially praises the code generation as 'gorgeous' but then immediately highlights a significant inconsistency: whether the C++ reference code is inside or outside the HTML output, depending on a 'thinking enabled' setting. This discrepancy suggests the output isn't as consistently 'gorgeous' as initially claimed, indicating a mismatch between the confident praise and the observed behavior.
Original quote: “definitely a gorgeous generation. Here with thinking enabled 4,000 tokens later, we have actually C++ reference code. So I asked it specifically provide HTML visualization and reference C++ code for both. So with thinking disabled, it had the reference code inside the HTML. With thinking enabled,…”
See the full analysis with timestamps →