How much FASTER is GLM 5.2 with MTP 💪 | Local AI TEST
Credibility score: 44/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "How much FASTER is GLM 5.2 with MTP 💪 | Local AI TEST" and rated it 44/100 for credibility (a BS score of 56/100 — mixed credibility), on 2026-06-25. Its weakest claim — "Speaker claims a peak of 29 tokens/second and considers it for the thumbnail." — scored 20/100 and was flagged as cherry-picked. 19 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 19 claims analyzed: 1 scored under 40, 18 between 40 and 69, and 0 at 70 or above.
Claims analyzed
Video opens with a highlight reel preview of speed metrics. — Just Vibes (50/100)
At 0:00
Just a quick-cut montage of numbers flashing by — setting the stage for a speed test, not making a specific claim yet. ⚡️
Why this score: The speaker is showing rapid-fire numbers (tokens per second) without context, which is typical for a video intro or highlight reel. It's meant to build anticipation for the main topic, not to be a verifiable claim itself. It's pure setup.
Original quote: “of the situation. Oh, 20 tokens, 13
tokens a second, 16, 20, 5, 20, actually
29 tokens a second. I flashed at 29
tokens a second. It flashed it. 24
tokens.”
Claiming MTP decoding sped up other models 'up to two times speed'. — Missing Context (45/100)
At 0:21
Says 'up to two times speed' for other models, but doesn't name a single one or give any specifics. 'Other models' is doing a lot of heavy lifting. 🤷♀️
Why this score: The speaker makes a general claim about MTP decoding's impact on 'other models' without providing any specific examples, benchmarks, or sources. This lack of detail makes it hard to verify the claim or understand the context of this 'two times speed' improvement. It's a broad statement used to set expectations for GLM 5.2.
Original quote: “that in other models has sped them up up
to two times speed. So, we're going to
see how much faster we can make GLM 5.2
with multi predictions.”
Explains MTP uses a tiny layer to predict what a massive model would, calling it 'asking a lot'. — Confidence Mismatch (45/100)
At 0:30
Describes the MTP layer as 'very very small' and the main model as 'half a terabyte gigabyte' — then says 'you're asking a lot' with no actual scale. That's a lot of hand-waving. 🤏
Why this score: The speaker uses vague, yet confident, language to describe the MTP layer's size ('very very small') and the main model's size ('half a terabyte gigabyte' — which is a weird combo of units, implying a massive scale without being precise). The conclusion 'you're asking a lot' is an emotional appeal to the difficulty rather than a technical explanation of the challenge. It's a confidence play without concrete metrics.
Original quote: “multi-time prediction does it, kind of it just has one layer that's very very small and it tries using that layer to predict what the main half a terabyte gigabyte model would predict. So, you're asking a lot.”
States higher quantized models run faster, but Q2/Q3 won't speed up with NTP. — Confidence Mismatch (45/100)
At 0:30
Claims higher quantization makes it faster, then immediately says lower quantization (Q2/Q3) won't speed up. Kinda contradictory, no? 🤔
Why this score: The speaker makes a general statement that 'it does run faster if you're using a higher quantized model,' which is a common understanding in model optimization. However, they immediately follow this with a specific caveat that 'if you're trying to use like a Q2 or a Q3, using NTP isn't going to speed it up.' This creates a slight mismatch in confidence because the initial broad claim of 'faster with higher quantization' is immediately limited without a clear explanation of why Q2/Q3 specifically fail to benefit from NTP, making the initial 'bonus' feel less universally applicable than implied.
Original quote: “Now, another bonus thing is it does run faster if you're using a higher quantized model. So, if you're trying to use like a Q2 or a Q3, using NTP isn't going to speed it up.”
Guaranteed speed above 20 tokens/second, then it immediately drops. — Confidence Mismatch (45/100)
At 2:30
Declared 'guaranteed' above 20 tokens/second, then the live meter immediately dipped below. The guarantee had a shelf life of like, 2 seconds 💀
Why this score: The speaker makes a strong, confident assertion ('guaranteed') about the performance, but the live data on screen (and their subsequent commentary) immediately shows the speed dropping below that threshold. This creates a mismatch between the stated confidence and the observed evidence.
Original quote: “we are above 20 tokens a second guaranteed.”
Claiming 'above 20 tokens a second guaranteed' while the live feed drops to 17. — Confidence Mismatch (45/100)
At 2:30
Said 'guaranteed above 20' then watched it immediately dip to 17. The 'guaranteed' part is doing some heavy lifting. 🤡
Why this score: The speaker confidently states 'we are above 20 tokens a second guaranteed,' but the live token indicator on screen immediately drops to 17 tokens/second. This creates a clear mismatch between the speaker's confident assertion and the real-time data being displayed, undermining the 'guaranteed' claim.
Original quote: “Lost 20 tokens a second. Okay, we are above 20 tokens a second guaranteed. Look, for now, second drops down to 17.”
Assuming coding tasks are easier for token prediction due to structure. 🤔 — Confidence Mismatch (45/100)
At 4:30
Assuming 'usually guesses the right tokens' for coding without showing the data. That's a vibe, not a fact. 🔮
Why this score: The speaker makes a generalization about coding tasks being easier for the model to predict tokens ('it usually guesses the right tokens') due to their structured nature. While this might be intuitively true, it's presented as a confident assertion without specific evidence or prior testing shown in this video to back it up, especially as they immediately say they've 'never actually tested this one out'.
Original quote: “Usually coding ones, cuz they're structured, it usually guesses the right tokens.”
Stating the MTP enabled result as 15.2 tokens/second. — Confidence Mismatch (45/100)
At 4:30
Said '15.2 tokens a second with MTP enabled' but the previous test was 15.8. Which one is it? 🤨
Why this score: The speaker previously stated the Eiffel Tower test (with MTP enabled) resulted in '15.8 tokens a second' at 5:22. Here, they state '15.2 tokens a second with MTP enabled' as if it's a general result, which contradicts the earlier specific test result for MTP enabled. This creates a mismatch in the reported numbers, making the claim less reliable.
Original quote: “And in total, it was 15.2 tokens a second with MTP enabled.”
Testing Flappy Bird Zero with MTP, claiming over 20 tokens/second but showing 15.4 and 15. — Confidence Mismatch (45/100)
At 6:30
Says 'over 20 tokens a second' then immediately shows 15.4 and 15. The numbers are doing their own thing 🤡
Why this score: The speaker claims a performance boost to 'over 20 tokens a second' with MTP, but the numbers displayed and stated immediately after are 15.4 and 15 tokens a second. This creates a mismatch between the stated expectation and the observed result, then they pivot to 'the thinking writing part it gets struggles with' to explain the discrepancy.
Original quote: “I'm just going to test that theory. I'm just going to do normal Flappy Bird zero with a baseline again of 15.4 tokens a second and with MTP with Flappy Bird zero that's over 20 tokens a second, 15 tokens a second. The thinking writing part it gets struggles with, but now”
Comparing MTP disabled vs. enabled, but the numbers are all over the place 🤡 — Confidence Mismatch (45/100)
At 8:30
Starts with 'nice stable 15.3' then immediately lists 11, 12, 14, 19, and ends with 'abysmal 13.4.' Which one is it, chief? 🤷♂️
Why this score: The speaker presents a 'nice stable' baseline, then immediately contradicts it with a rapid-fire list of wildly fluctuating numbers, concluding with a completely different 'abysmal' result. The confidence in the initial 'stable' figure doesn't match the subsequent data presented.
Original quote: “So, by default MTP disabled you get a nice stable 15.3 tokens a second. Now, the most it could write was 200 tokens, so this is going to be a good challenge for MTP, but oh my 11 tokens a second, 12 tokens a second, 14 tokens a second, 19 tokens a second. It's not doing 10 tokens a second for the…”
Predicting MTP will be 'bonkersly bad' at creative writing, then showing it struggle. — Confidence Mismatch (45/100)
At 8:30
Called it 'bonkersly bad' before the test, then showed it underperform. The prediction was a bit of a self-fulfilling prophecy, no? 🔮
Why this score: The speaker sets an expectation of poor performance for MTP in creative writing ('bonkersly bad') and then presents results that confirm this expectation, even though the initial 'bonkersly bad' is a subjective, unquantified prediction. It's like saying 'this car is gonna suck' then showing it lose a race. The confidence precedes the data, shaping the narrative.
Original quote: “So, this is kind of like creative writing and I'm guessing MTP is going to get creative writing completely bonkersly bad. So, write a poem about GLM. So, by default MTP disabled you get a nice stable 15.3 tokens a second. Now, the most it could write was 200 tokens, so this is going to be a good…”
Downplaying the cost of 5 gigs of memory with a shrug 🤷♂️ — Loaded Language (45/100)
At 10:30
Dismissing 5 gigs as 'probably not worth it' is a bit of an understatement for many users. That's a lot of RAM for a 'maybe' speed boost. 💾
Why this score: The speaker uses casual language ('probably not worth it') to downplay the significant memory cost (5 GB) for a potentially minor performance gain. This frames the trade-off as less impactful than it might be for users with limited resources, making the 'potential' seem more appealing.
Original quote: “Probably not worth the 5 gigs of memory,”
Downplays the cost of 5 gigs of memory with a shrug 🤷♂️ — Loaded Language (45/100)
At 10:30
Dismissing 5 gigs of memory as 'probably not worth it' is a casual way to frame a significant resource cost. 💾
Why this score: The speaker uses 'probably not worth it' to minimize the impact of 5GB of memory usage, making it sound like a minor consideration rather than a potentially hefty requirement for many users. It's a soft dismissal of a real cost.
Original quote: “Probably not worth the 5 gigs of memory,”
Admits MDP isn't performing well, despite improving accuracy from 10% to 70%. — Confidence Mismatch (45/100)
At 12:30
Went from 'Whoa' at 50% to 'not doing too well' at 70%. The goalposts are moving faster than the tokens. 🥅💨
Why this score: The speaker expresses initial excitement about improving prediction accuracy from 10% to 50% and then to 70%. However, immediately after reaching 70%, they state it's 'not doing too well.' This creates a mismatch between the perceived improvement and the current assessment of performance, suggesting the initial 'Whoa' was a bit premature or the expectations shifted.
Original quote: “14.9 tokens a second. Now with MDP, because this is something that it doesn't know, or maybe the implementation just needs improving. You know, when I started the implementation, it was getting a prediction accuracy of 10% and then I got it to 50% and I was like "Whoa." And now we got up to 70 and…”
Claims they're 'gaming' with MDP disabled, implying a performance boost. — Loaded Language (45/100)
At 12:30
Using 'gaming' to describe performance with MDP off — sounds like a win, but it's just the baseline. 🎮
Why this score: The term 'gaming' here is loaded, suggesting an advantage or optimization when it's simply the default state before introducing the new tech. It sets a positive tone for the 'disabled' state, which is then compared to the new feature's performance.
Original quote: “So, we're already gaming with MDP disabled.”
The numbers are all over the place, and the speaker attributes it to 'memory clearing' 🤷♂️ — Confidence Mismatch (45/100)
At 14:30
The token per second numbers are jumping around, and the explanation is 'maybe my memory needed clearing.' That's not exactly a scientific variable control, chief 🤡
Why this score: The speaker presents multiple, inconsistent token-per-second results (15, 16.4, 16.5, 18) for the same Q4 test, then vaguely attributes the discrepancy to 'memory clearing.' This undermines confidence in the consistency of the testing methodology and the results presented.
Original quote: “ended up the show with, although it's kind of struggling when it comes to the textual summarization task, 15 tokens a second. Yeah, we ended up with 16.4 tokens a second, which is actually if you look at it, somehow it was faster than that one. I think my I think I just need to do that test one…”
Declares four steps are definitely slower than three in this test. — Confidence Mismatch (45/100)
At 16:30
Says 'definitely slower' based on one run. That's a strong word for a single data point. 😬
Why this score: The speaker uses the word 'definitely' after what appears to be a single test run. While the observed result might show four being slower, claiming it's 'definitely' slower without multiple runs or statistical analysis is an overstatement of confidence.
Original quote: “So four is definitely slower than three in this situation.”
Speaker claims a peak of 29 tokens/second and considers it for the thumbnail. — Cherry-Picked (20/100)
At 16:30
Saw a peak of 29 for a second, but the average was 18. Cherry-picking the high point for the thumbnail, classic move. 📈🖼️
Why this score: The speaker is excited about a momentary peak performance of 29 tokens/second, even though the sustained or average performance for that configuration (five tokens) was 18 tokens/second. Suggesting this peak number for the thumbnail is a classic example of cherry-picking the most impressive, but not necessarily representative, data point to create a more impactful impression.
Original quote: “I'm getting goosebumps right now cuz I saw 29 tokens a second with GLM 5.2. I might just make that the the the thumbnail. 29 tokens a second.”
Rapidly listing various token-per-second results, ending with a specific 'flashed' speed. — Confidence Mismatch (45/100)
At 18:30
Just rattling off numbers with no context or explanation for the fluctuations. It's a speed run of stats. 💨
Why this score: The speaker is listing several different token-per-second figures in quick succession without explaining the conditions or variations that led to each number. While they end on a specific 'flashed at 29 tokens a second,' the preceding rapid-fire, uncontextualized numbers make it hard to track or verify the overall performance claims.
Original quote: “Oh, the situation. Oh, 20 tokens, 13 tokens a second, 16, 25, 20, actually 29 tokens a second. I flashed at 29 tokens a second. I flashed it. 24 tokens”
See the full analysis with timestamps →