Kimi K3 Open Source AI Beats GPT & Fable? 🤯 (Coding FF7, Red Dead & MLX)
Credibility score: 44/100 — Mixed Credibility. Several questionable claims detected. Watch with healthy skepticism.
BSmeter analyzed "Kimi K3 Open Source AI Beats GPT & Fable? 🤯 (Coding FF7, Red Dead & MLX)" and rated it 44/100 for credibility (a BS score of 56/100 — mixed credibility), on 2026-07-18. Its weakest claim — "Made full FF7, RDR2, and GTA5 remakes with one model" — scored 20/100 and was flagged as confidence mismatch. 41 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 41 claims analyzed: 3 scored under 40, 38 between 40 and 69, and 0 at 70 or above.
Claims analyzed
2.8 trillion param Kimi K3 beats GPT & Fable — zero model shown — Confidence Mismatch (45/100)
At 0:16
Drops a 2.8-trillion claim and leaderboard before any model or benchmark loads. Bold. Stupid, but bold. 😈
Why this score: Speaker states a specific parameter count and implies superiority over named models, yet the transcript offers no source, benchmark link, or even the actual leaderboard graphic. The confidence arrives before the evidence.
Original quote: “Hey guys, welcome to the show. Today we're checking out Kimmy K3 2.8 trillion parameter model open weight. This is fable GPT sol intelligence in the open realm arena. Look at these scores over here. BT, GPT, SOL, and Fable 5 in some”
Drops 2.8-trillion-param model with no source or benchmark link — Confidence Mismatch (45/100)
At 0:16
Announces 2.8 trillion parameters like it’s settled fact — zero receipts, zero paper, zero link. 😈
Why this score: The number is thrown out with full presenter confidence, but the transcript never names the origin, the benchmark suite, or any verifiable source. That’s textbook confidence mismatch: certainty without the data that would earn it.
Original quote: “Hey guys, welcome to the show. Today we're checking out Kimmy K3 2.8 trillion parameter model open weight. This is fable GPT sol intelligence in the open realm arena. Look at these scores over here. BT, GPT, SOL, and Fable 5 in some”
Kimi K3 never stops until 100% happy — built-in harness claim — Confidence Mismatch (45/100)
At 0:48
Says the model has a 'builtin harness' that keeps going until 100% — names no code or test that proves it.
Why this score: The speaker presents the self-looping behavior as a special feature of this model, but gives zero evidence that it's actually built-in versus just how the prompt or inference is set up. Classic 'I observed this, therefore it's engineered that way' leap without receipts.
Original quote: “it’s making me very, very lazy. I think with this model, it’s got a builtin harness. So rather than ending the turn with a finished job, it goes ahead and double checks everything, downloads files, continues on. So normally you’d have like a third party running a model, this one here doesn’t stop.…”
Made full FF7, RDR2, and GTA5 remakes with one model — Confidence Mismatch (20/100)
At 1:13
Says 'made' like he shipped finished games — these are tiny demos at best, not remakes.
Why this score: Real remakes take years and teams. He's running a local LLM that generated some WebGL scenes. The gap between 'made a remake' and 'prompted some assets' is massive.
Original quote: “I made Final Fantasy 7 remake. I made Red Dead Redemption 2 remake. I made GTA 5. I made a human face and I even got it to to port a transformers model in MLX”
Calls built-in harness the 'main innovation' with zero comparison — Confidence Mismatch (45/100)
At 2:35
Labels it the core breakthrough like that's obvious — no benchmarks, no prior models mentioned. 😈
Why this score: Speaker presents the 'harness' as the standout feature without showing what other models already do or why this version is meaningfully different. It's confident framing without receipts.
Original quote: “the main innovation of this one is the built-in harness where it just doesn't stop until it's happy”
Lists basic edits as if they're new frontier capabilities — Missing Context (45/100)
At 2:52
Acts like skin tone and age sliders are mind-blowing — most image models have had this for years. 😈
Why this score: The speaker presents standard prompting controls as novel potential without acknowledging that these features are table stakes across diffusion and multimodal models already.
Original quote: “You can change the skin tone. You can even change the age. You can add in a bit stubble”
Jumps from 'potentially' to asking for full AAA browser game in one breath — Confidence Mismatch (45/100)
At 3:13
Slides from 'maybe' to demanding AAA graphics and proper models with zero evidence it can deliver. 😈
Why this score: The prompt escalates instantly from a casual suggestion to expecting a full-scale, high-fidelity game. The model correctly pushes back, but the framing still treats it as a reasonable ask.
Original quote: “potentially we can actually make GTA 6 ourselves, but can you make GTA V that runs in a web browser? AAA graphics and proper models”
Jumps from prompting an AI to 'make GTA 6 ourselves' — Confidence Mismatch (45/100)
At 3:14
Said 'make GTA 6 ourselves' like the prompt is the hard part. Bold. Stupid, but bold.
Why this score: The speaker treats the creative leap as solved once the prompt is written, ignoring the actual engineering gap between a text prompt and shipping a full AAA game.
Original quote: “potentially we can actually make GTA 6 ourselves”
Model correctly flags the absurd ask — speaker still calls the attempt 'doing some stuff' — Missing Context (45/100)
At 3:32
Model tells the truth about scope and budget — speaker shrugs it off and keeps going anyway. 😈
Why this score: The AI gives a clear, accurate reality check on why this can't happen. The speaker acknowledges it yet immediately downplays it as the model still 'doing some stuff,' softening the failure.
Original quote: “GTA 5 took a thousand person studio years and a nine figure budget. You're not going to do it with me, bro”
Bugs exist but model is still great — Missing Context (45/100)
At 4:26
Admits collisions are broken — then immediately pivots to 'so much good' without showing the fix.
Why this score: The visual failure is waved away with optimism. No evidence the reprompt actually resolves it; the claim rests on hope rather than demonstrated correction.
Original quote: “I can't say that. I collide with the buildings, but they don't collide. Collide.”
Calls current output 'not best' while promising quality with tweaks — Missing Context (45/100)
At 4:46
Promises 'quality game' with extra turns — never shows what that actually produces. 💀
Why this score: Keeps the bar floating just out of reach. Shows buggy demo, then immediately frames it as fixable with more prompting, so the viewer can't judge the current ceiling.
Original quote: “So, with proper models and maybe with a few more turns, we could actually probably get a quality game. But this isn't the best demonstration.”
Prompt upgrade equals better game — Missing Context (45/100)
At 4:58
Calls the prompt 'more ambitious' while the actual output still shows placeholder buildings and broken physics.
Why this score: Ambition in the text prompt is treated as proof of progress. The visuals don't reflect the jump; only the request got bigger.
Original quote: “This is Red Dead Redemption 2. And I'm going to show you my prompt. My prompt is getting more ambitious.”
Audio generation is now a 'skill' — Loaded Language (45/100)
At 5:25
Labels basic TTS as a learned 'skill' like the model gained a new ability rather than just calling an API.
Why this score: The framing turns routine tool use into evidence of emergent capability. No distinction made between native generation and external TTS service.
Original quote: “It's getting some skills and one of the skills is audio generation. So, we're going to make some AI generated audio TTS”
Calls it a 'vision model' that 'sees and understands' — Loaded Language (45/100)
At 5:44
'Understands' is doing heavy lifting — it's generating previews, not proving comprehension. 🔥
Why this score: Equates visible output with actual understanding. The screenshots are just the model spitting back what it was told to make; no evidence of deeper scene comprehension.
Original quote: “The great thing about Kimmy is a vision model. So, you can actually see and understand what it's seeing.”
Attributes 'excitement' to the model about its own work — Loaded Language (45/100)
At 5:58
Model isn't excited — it's role-playing the prompt. Anthropomorphism doing the selling. 😈
Why this score: The prompt told it to act excited. Calling the output 'genuine excitement' is just reading the script back as personality.
Original quote: “It's actually getting excited with itself and it work, which it should be, which is great.”
5-hour FF7 + MLX port — no receipts, just time claim — Confidence Mismatch (45/100)
At 8:04
Drops '5 hours' like it's measured — zero logs, zero timestamps, zero proof it wasn't 20. 💀
Why this score: Time spent is presented as fact with zero supporting detail. No workflow, no error logs, no comparison to other attempts. The number does the heavy lifting while the evidence stays off-screen.
Original quote: “I got one more, actually two more still to show you. I got Final Fantasy 7 as well as porting to MLX. And that one took maybe 5 hours to do.”
Calls it AAA graphics while showing early dark prototype — Confidence Mismatch (45/100)
At 8:36
Labels it 'AAA graphics' while the actual screenshots are still very dark and unfinished.
Why this score: The term 'AAA' implies polished, high-budget quality. The model just generated blackish textures and the speaker immediately had to ask it to fix them — so the ambition claim is running way ahead of the actual output shown.
Original quote: “this is original Final Fantasy 7 style game clone but with AAA graphics. I'm being ambitious here.”
Claims it's AAA graphics on an FF7 clone — confidence without proof — Confidence Mismatch (45/100)
At 8:36
Calls it AAA graphics while describing a hobbyist AI clone. That's a stretch with zero comparison data.
Why this score: Speaker labels the output AAA-level while showing dark, low-detail textures and admitting it needs fixes. The gap between the label and what's actually displayed is the move.
Original quote: “So this is original Final Fantasy 7 style game clone but with AAA graphics. I'm being ambitious here.”
Assumes the AI is 'happy' with its own dark screenshots — anthropomorphism as evidence — Just Vibes (50/100)
At 8:50
Treating the AI saying 'done' like it has taste. Cute bit, but it's just code stopping.
Why this score: The speaker projects satisfaction onto the model because it stopped generating. That's entertainment framing, not a technical claim.
Original quote: “It takes screenshots here. So let's see what kind of screenshots it saw from itself. That's the intro scenes. It must be happy with that.”
Prompted the AI to fix dark textures after it already called itself done — iterative coaxing presented as success — Missing Context (45/100)
At 9:06
The model didn't nail it first try. User had to step in and correct it, then showed the improved version.
Why this score: The narrative frames the AI as autonomously completing a game, but the actual process required human intervention to fix the lighting. That's the omitted step.
Original quote: “Problem is it was very very blackish the textures and all that kind of stuff. So I said hey it's a bit black. Fix this stuff.”
Presents MLX Linux attempt as impressive without showing results — Missing Context (45/100)
At 9:26
Mentions it downloaded MLX on Linux but leaves out whether anything actually worked or just failed.
Why this score: The sentence stops at 'had to run it in CPU mode' — no follow-up on speed, errors, or output quality. The implication is that the model is doing advanced stuff, but the viewer gets zero evidence it succeeded.
Original quote: “It actually even downloaded MLX and tried running it on Linux and then it had to run it in CPU mode.”
One prompt created full FF7 scene — confidence mismatch — Confidence Mismatch (45/100)
At 11:16
Calls it 'just one prompt' like that explains the entire result. Zero detail on what was actually fed in.
Why this score: The phrasing downplays the massive amount of prior work (models, training data, fine-tuning, prompting craft) and sells the illusion that anyone could do this with a single casual line. Classic move to make the tech sound more magical than it is.
Original quote: “Just one prompt made this.”
Tifa or Eric romance choice — false dilemma — False Dilemma (20/100)
At 11:35
Presents only two romance options when the original game and mods have way more paths.
Why this score: The model is being praised for 'staying true,' yet the speaker immediately reduces the choice to Tifa vs. a misspelled stand-in. That shrinks the actual narrative scope and makes the demo look more faithful than the limited output actually is.
Original quote: “we're going to have to pick who we're going to go out with. Is it going to be Tifa or is it going to be Eric?”
Gemini and Claude failed, Kimi succeeded — anonymous authority on other models — Anonymous Authority (45/100)
At 13:52
Names Gemini Pro and Claude Sonnet Max failing the same prompt with zero logs or output shown. 💀
Why this score: Speaker drops two specific model names as failed benchmarks, yet provides no evidence of what prompts were used, what errors occurred, or any side-by-side comparison. Classic 'I tried the big names and they sucked' move without receipts — the authority is in the names alone.
Original quote: “I even did this exact same prompt with Gemini Pro and Claude Sonnet Max and they couldn't figure out what it is. So, I let Kimmy do its thing and it went ahead and read the files, tried figuring it out, tried going what's going on and even tried downloading stuff on the internet and it said it was…”
Calls the fix 'sneaky' then walks it back — Loaded Language (45/100)
At 14:30
Labels the code change 'sneaky' for drama — then immediately says he doesn't know if it is.
Why this score: Loaded emotional word doing the heavy lifting before he undercuts it with 'I'll let you decide.' Classic volume game: strong claim, instant retreat.
Original quote: “this is actually really sneaky. I don't know if it's sneaky. I'll let you guys decide.”
LLM diagnosed quantization as the bug — no proof it actually checked — Anonymous Authority (45/100)
At 14:35
AI said it 'checked everything' — zero receipts on what it actually looked at. Classic anonymous authority move.
Why this score: Speaker presents the model's internal diagnosis as authoritative without showing any logs, diffs, or verification steps. The claim rests entirely on 'it told me' rather than demonstrable evidence.
Original quote: “It told me that it went ahead and checked everything out and it says that the issue is that it's free bits. It's uh too strong quantization.”
Claims the real bug was something else entirely — Missing Context (45/100)
At 14:46
Says 'this was the bug' like it's obvious — never shows what the original prompt or expected behavior actually was.
Why this score: The speaker is redefining what counts as the 'real' bug after the fact. Without the original task description, we can't tell if adding the comma zero was the intended fix or just a lucky accident.
Original quote: “Now, if we look at what he actually did versus what he actually said. So, here is one of the edits it's made and it added in a little predicate to make sure that the ln head doesn't get quantized. That's good. Generally, that wasn't the bug. This was actually the bug.”
Claims the real bug was elsewhere — but the 'good' edit gets zero scrutiny — Missing Context (45/100)
At 14:46
Calls the LN head fix 'good' then immediately dismisses it without showing why it wasn't the actual problem.
Why this score: Speaker sets up a contrast between stated and actual changes but provides no evidence that the LN head edit was irrelevant. The dismissal happens without data or further testing shown.
Original quote: “Now, if we look at what he actually did versus what he actually said. So, here is one of the edits it's made and it added in a little predicate to make sure that the ln head doesn't get quantized. That's good. Generally, that wasn't the bug.”
Missing comma zero declared the sole bug — no evidence it's the only one — Confidence Mismatch (45/100)
At 15:01
Calls the missing comma 'the bug' with total certainty while admitting the model found it by accident.
Why this score: Speaker asserts one simple syntax error as the definitive root cause. No testing shown ruling out other issues, and the fix was discovered serendipitously rather than through systematic debugging.
Original quote: “This was actually the bug. He actually did the bug, but he didn't know that it fixed the bug. This was the bug. It was so simple. When we were calling next, we had the self. So, this line of this code block was working fine except this next it never had a comma zero.”
Calls the missing comma 'so simple' after AI found it — Confidence Mismatch (45/100)
At 15:05
Calls the fix trivial only after the model spotted it. Before that, it was apparently invisible to him.
Why this score: Retroactive 'it was obvious' energy. The bug sat there until the AI caught it, yet now it's framed as something anyone would have seen immediately.
Original quote: “It was so simple. When we were calling next, we had the self. So, this line of this code block was working fine except this next it never had a comma zero. So it added in a comma zero and that fixed the gibberishness”
2.5 tokens/sec called 'very very good' then immediately a problem — confidence flip — Volume Game (45/100)
At 15:20
Praises the output as 'very very good' then instantly pivots to speed as the new crisis.
Why this score: Speaker loudly celebrates functionality, then quietly undercuts it with performance numbers in the same breath. The positive claim gets the volume; the limitation gets the quiet follow-up.
Original quote: “So when you now go ahead and say hello it actually responds and it talks to you and it was very very good. But there was another problem. It was running at 2.5 tokens a second.”
Model couldn't follow 'don't over-optimize' instruction — presented as model's failure — Missing Context (45/100)
At 15:45
Blames the model for not obeying the instruction without showing what the actual spec said.
Why this score: Speaker frames the model's inability to comply as a limitation, yet never quotes or links the original specification. The claim that it 'couldn't' do it rests on an unshown document.
Original quote: “So I said, "Hey, don't do that because um it's not part of the the original specification. Let's not optimize overoptimize anything." So I gave it that and it couldn't wasn't able to do it.”
Found external GitHub that already works — no comparison shown on screen — Missing Context (45/100)
At 15:53
Claims the other repo is better but never displays the actual code differences during the segment.
Why this score: Speaker references an external solution as superior without presenting side-by-side diffs or benchmarks in the video. The audience is asked to accept the superiority on description alone.
Original quote: “But then check this out. I found this some dude pipe network. They already had a sort of port of it running in MLX natively. So I said, "Hey, this is the GitHub. Go ahead. Can you take a look at it and just see what our code is missing compared to its code?"”
Sets 20 tokens/sec as the bar based on a random GitHub port — Confidence Mismatch (45/100)
At 15:53
Picks 20 t/s as the success threshold with zero explanation of why that number.
Why this score: The target speed appears pulled from thin air. No benchmark, no hardware spec, just 'I'm expecting 20.' The number does the work of sounding rigorous without any supporting logic.
Original quote: “I found this some dude pipe network. They already had a sort of port of it running in MLX natively. So I said, 'Hey, this is the GitHub. Go ahead. Can you take a look at it and just see what our code is missing compared to its code?' And I said, I'm expecting it to be at least 20 tokens a second.”
Model found the attention difference — no verification that this was actually the bottleneck — Confidence Mismatch (45/100)
At 16:07
Presents the fused attention diagnosis as fact while admitting the 20 t/s target was just a guess.
Why this score: Speaker states the model's conclusion with certainty even though the original performance expectation (20 tokens/sec) was arbitrary. No independent profiling shown to confirm attention was the limiter.
Original quote: “And I said, I'm expecting it to be at least 20 tokens a second. It went ahead and looked through and it found a difference between the code we had wrote and the code it had wrote. And it said that I was using it said I was using manual attention rather than the fused STPA.”
Model doubles down with '100%' certainty on wrong diagnosis — Confidence Mismatch (20/100)
At 17:40
Says 'completely 100%' while being dead wrong — confidence without evidence, the oldest trick in the book.
Why this score: The model escalated from suggestion to absolute certainty in one line even though the speaker later reveals the actual issue was a simple code bug. That 100% is pure theater; it has no data backing the claim and ignores the user's direct correction.
Original quote: “No, no, it's completely 100% to do with the quantization.”
Calls Kimi K3 "really, really good" on game coding with zero metrics — Confidence Mismatch (45/100)
At 18:30
Zero benchmarks, zero comparison, just vibes — confidence without receipts 💀
Why this score: Speaker jumps from "99% automated" to blanket praise without showing any actual performance numbers, error rates, or side-by-side results against GPT/Fable. The excitement is real; the evidence is missing.
Original quote: “So that is exciting. And even with these games that we made, it was really, really good. Good good work.”
Kimi K3 might have built-in harness vs. added magic — pure speculation dressed as insight — Confidence Mismatch (45/100)
At 18:55
Admits zero idea how it runs, then confidently suspects built-in harness like it's obvious. Bold guess, zero receipts.
Why this score: Speaker openly states they have no idea, yet pivots straight to a specific technical suspicion about the model's internal behavior without any evidence shown. Classic move of presenting an untested hunch as the most likely explanation.
Original quote: “I have no idea how we're going to run it. I'm very very excited to see how it'll actually even run. Like, is does it actually have the harness built in like I suspect? Cuz that's what seems to be doing. It doesn't end its turn until it's happy with it. Or is that just some added magic that they've…”
Dramatic training montage story with no evidence shown — Just Vibes (50/100)
At 20:47
Spun a whole Rocky training arc off one AI-generated boxing game — zero receipts on the five weeks.
Why this score: Pure entertainment beat — he's selling the struggle story for the bit, not claiming it's real. The TTS line later even undercuts it.
Original quote: “We did it, Adrian. So, five weeks, 5 weeks, 4:00 a.m. alarms, cold piers, and a city that learned his name. Kingston hits the hits like a truck, but he drives like one, too.”
One prompt claim + crowd size flex with no proof — Confidence Mismatch (45/100)
At 21:32
Asks if you expected this from one prompt while showing nothing but the running game — crowd size is just asserted.
Why this score: The 'one prompt' line is doing the heavy lifting to impress, but the only thing demonstrated is a Temple Run clone. The 14k audience is dropped like a flex with zero verification on screen.
Original quote: “Did you expect this with one prompt? So, 14,000 people. Look at him. Look. Look. Look at his knees.”
Winning the fight by mashing keys while admitting he doesn't know what's happening — Just Vibes (50/100)
At 22:03
Won the boxing match by button-mashing and openly said he has no idea how the game works — peak chaos.
Why this score: He's not pretending this is impressive AI output; he's roasting his own clueless victory. The honesty makes it funny instead of a claim.
Original quote: “Oh, I'm winning. I'm buy mashing and I'm winning. I don't know. Yo, I knocked him down. Okay, I'm going to win this game. I don't know exactly what's going on. No punches are being thrown. I'm just smashing out the keys and I've knocked him down again.”
See the full analysis with timestamps →