Llama.cpp Just Merged MTP And You Should Be Using It.
Credibility score: 66/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
BSmeter analyzed "Llama.cpp Just Merged MTP And You Should Be Using It." and rated it 66/100 for credibility (a BS score of 34/100 — mostly credible), on 2026-05-18. Its weakest claim — "MTP was previously missing from Llama.cpp, and they will now explain what it is." — scored 50/100 and was flagged as just vibes. 12 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 12 claims analyzed: 0 scored under 40, 7 between 40 and 69, and 5 at 70 or above.
Claims analyzed
Llama.cpp now has MTP, promising 25%+ speed boost without any downsides. — Solid (80/100)
At 0:00
Finally! They've merged MTP into Llama.cpp and are claiming zero trade-offs? I'm skeptical but intrigued 🤔✅.
Why this score: MTP Integration Confirmed: The web context confirms the recent merge of Multi-Token Prediction (MTP) support into Llama.cpp.
- Performance Claim: The speaker claims a '25% or even more increase in tokens.' This aligns with general expectations for MTP implementation, which often yields 1.5x to 2x gains.
- Hardware Agnostic: This is standard for modern Llama.cpp builds, so that part checks out.
Original quote: “Hey everybody, Timothy Karen, local AI enthusiast, and today I want to talk about how finally something that has been around for actually a long time is now finally present in the most popular tool for running models locally, Llama CPP. And this should actually give you like a 25% or even more…”
Software advancements like 1-bit models and TurboQuat can enable large models (200B parameters) to run locally. — OK (65/100)
At 0:50
Mix of real tech and a *very* optimistic outlook on local LLM limits 🧐
Why this score: While 1-bit quantization and context cache improvements like TurboQuant are valid approaches to optimize local LLM inference, the speaker's claim that *software alone* can unlock "on device intelligence" for "200 billion parameter models" without significant new hardware advances is an exaggeration. Hardware limitations remain substantial.
Original quote: “I have this crazy idea that even if no new local models were even invented like like just what we have is what we have. There is still so much juice to squeeze at the software level that we could unlock on device intelligence still like this would still be possible. These super large you know 200…”
Timothy Carbat is the founder and creator of Anything LLM. — Personal Story (75/100)
At 2:19
Okay, quick intro. Got it. 👍
Why this score: The speaker is introducing himself and his project, which falls under personal testimony. Without external verification, such claims are typically considered personal stories. However, if 'Anything LLM' is a known project and he is associated, it's a solid personal story.
Original quote: “For those of you who don't know me, quick intro. My name is Timothy Carbat, founder and creator of Anything LLM.”
Llama.cpp received Multi-Token Prediction (MTP) support two days ago. — Verified (90/100)
At 2:19
✅
Why this score: Multi-Token Prediction (MTP) support was recently merged into Llama.cpp in May 2026. Given the current date of May 18, 2026, the claim of "two days ago" (May 16, 2026) is consistent with official release information and community discussions around the MTP merge.
Original quote: “I want to bring your inference local, but obviously I don't want to sacrifice the output or the speed that you're probably used to. So, two days ago, we finally got MTP support in Llama CPP.”
MTP was previously missing from Llama.cpp, and they will now explain what it is. — Just Vibes (50/100)
At 2:30
The 'high horse' comment felt a little smug, but the promise of a crash course is tempting. 😏🔥
Why this score: Commentary on the state of Llama.cpp users—it implies that using it *before* MTP was peak performance bragging rights. The speaker sets up the entire segment to justify why this new feature matters.
Original quote: “So, you can just sit on your high horse and be happy, I suppose. And if you're on Llama CPP, then we finally got support. Now, why was MTP missing? What is MTP? These are all the things that I'm going to go over super super crash course on this stuff.”
MOE models (like 122 active 10B) might not show the same MTP speed gains as others. — OK (65/100)
At 4:30
They're right to caution us about MOEs. It’s not a universal silver bullet! 🙄⚠️
Why this score: Caveat is necessary because the architecture changes how MTP benefits are realized. - While MTP works across architectures, MoE models have inherent routing overhead that can temper speedups compared to dense models. - This isn't outright false, but it oversells the universal applicability of the 2x/3x claim.
Original quote: “Word of caution though, MOE models or the mixture of expert models like for example this 122 active 10B, you may not see the performance improvements everybody else is getting using ane.”
MTP isn't perfect ('silver bullet'), but it works fine even when handling vision inputs. — Solid (75/100)
At 6:30
So it's not magic, but it's definitely *better* than nothing. ✨👌 — This addresses the common 'it only works on text' assumption.
Why this score: Nuance is key here; they aren't overselling it.
- Vision Compatibility: Confirms MTP extends beyond pure token generation.
- Caveat: The speaker immediately tempers this by saying vision input will cause a 'slight hit' to prompt processing speed.
Original quote: “MTP is not a silver bullet as I've mentioned but it does work with vision so if you have vision models or vision input this will still function”
Unsloth models, like the 9B, now offer distinct MTP GGUF files in addition to standard GGUF versions. — Verified (90/100)
At 8:30
Yep, dedicated MTP model files are a real thing now. ✅
Why this score: The integration of Multi-Token Prediction (MTP) into Llama.cpp has led to the release of specific MTP-enabled GGUF model files by developers like Unsloth AI, alongside existing non-MTP versions.
Original quote: “you'll see there are now two files. There's the 9B GGUF, which is probably what you have downloaded uh or of whatever variant you are running right now. But then there's now this 9B MTP GGUF.”
The speaker is about to run a performance comparison/benchmark. — Just Vibes (50/100)
At 10:30
Alright, the hype train is leaving the station! — Let's see if this benchmark actually proves the point. 🚂💨
Why this score: Transition statement. This isn't a claim itself, but it sets up the core evidence for the entire segment. It signals that the subjective 'urge to use it' is about to be backed by hard numbers.
Original quote: “So now let's look at a very quick benchmark.”
Setting `draft_max` to 6 will dramatically show performance degradation. — Just Vibes (50/100)
At 12:30
They're about to prove the trade-off point! — Six tokens ahead? That’s asking for trouble. 😬🔥
Why this score: Setup for Proof. This isn't a claim itself, but it sets up the crucial demonstration: that increasing n past optimal points introduces overhead and slows things down.
Original quote: “And to just really illustrate the problem here, let's turn draft max to six. And so, I'm going to load this up and we're going to rerun that same exact prompt.”
Using MTP at N=6 results in 30 tokens/sec, which is slow without accuracy gain. — Just Vibes (50/100)
At 14:30
Said getting slower just for the hell of it? 😂 That's peak demo commentary. — But hey, at least we know N=6 isn't always optimal.
Why this score: Performance observation is clear here: The speaker notes that running MTP with a specific configuration (N=6) yields 30 tokens/sec without any corresponding accuracy boost. This suggests the overhead of MTP in this specific test case outweighs the benefit, making it inefficient for raw speed gains.
Original quote: “n equal to 6. I'm sitting at about 30 tokens a second. Keep in mind this is not the trade-off you want with MTP because there is no accuracy change. So, I am just getting slower tokens just for the hell of it, i guess. Like, it really doesn't make any sense to do it to use MTP and have a slower…”
Speaker shares their opinion on AI's role as a tool, not AGI, and questions the sustainability of cloud models. — Opinion (50/100)
At 16:30
Hot take alert! Dismissing AGI and sci-fi futures while championing local AI. Groundbreaking stuff. 🙄
Why this score: The speaker is expressing a series of personal beliefs and observations about the nature of AI, its current capabilities (not AGI), its future (not sci-fi levels), and the economic/ecological aspects of cloud computing. These are subjective perspectives rather than verifiable factual claims. While many share these views, they remain within the realm of *opinion*.
Original quote: “cloud model paradigm that exists right now is economically or ecologically sustainable in so many dimensions. But it's clear that AI is a tool and it is useful. It's not AGI. To me, AI is just a tool. It's a thing I use to help me with just general tasks. It's not going to be, you know, Jarvis.…”
See the full analysis with timestamps →