Mistral Large 4 vs Kimi K3 vs GLM 5.3

Author: Data Science in your pocket · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Mistral Large 4 vs Kimi K3 vs GLM 5.3 · Channel: Data Science in your pocket · 04:47 · Watch on YouTube

Topics: AI evaluation and benchmarks, Model inference and optimization, AI models, Local and open-source models

Summary

The video compares Mistral Large 4, Kimi K3, and GLM 5.3, arguing that benchmark winners depend on the task, the evaluation method, and the costs involved. It explains how mixture-of-experts models use only a portion of their total parameters per token, and notes that the models’ million-token context claims do not translate equally into measured performance. Kimi is presented as strongest for coding and agentic workloads, GLM for broad quality and value, and Mistral for low-cost API use and EU hosting. For self-hosting, the video says GLM 5.3 Flash is the only practical option among the models discussed, while emphasizing that open weights alone do not make a model accessible to hobbyists.

Key points

The comparison covers Mistral Large 4, Kimi K3 from Moonshot, and GLM 5.3 from Z.A.I.
All three are mixture-of-experts models, so total parameters represent what is stored while active parameters indicate what is used per token.
Although all three claim a one-million-token context, evaluators measured Mistral Large 4 closer to 512K; Kimi K3 is described as engineered for long context.
GLM 5.3 leads the cited broad-quality index, while Kimi K3 is presented as stronger for coding and agentic tasks.
The video cites GLM’s lower cost per coding task compared with Kimi, while noting that Kimi’s API pricing is not especially cheap.
The benchmark comparisons have caveats, including company-run evaluations, different baselines, and safety results affected by refusals.
The flagship models are described as impractical to self-host; GLM 5.3 Flash is the exception, with quantized weights said to fit a 128 GB Mac.
The video’s monthly API cost estimates are $289 for Mistral, $600 for GLM, and $1,620 for Kimi at 10 million tokens per day.

Chapters

Start	Chapter	Summary
00:00	The contenders and their claims	The video introduces the three models, their headline sizes, and the conflicting benchmark claims that motivate the comparison.
01:05	Active parameters, context, and benchmark strengths	It explains active parameters and context limits, then compares the models across quality, coding, agentic tasks, and specialized uses.
02:42	Benchmark caveats and self-hosting constraints	The narrator flags limitations in the benchmark claims and explains why the three flagship models are difficult to run independently.
03:44	GLM Flash and monthly costs	The video identifies GLM 5.3 Flash as the runnable option, notes its reported evaluation limitation, and compares estimated monthly API bills.
04:15	Verdict and the limits of open weights	The narrator gives recommendations by use case and concludes that access to open weights does not guarantee that hobbyists can run a model.

Quotes

Total parameters are the rent. Active parameters are the speed. — 01:05

Nobody lied here. They just picked the baseline, the harness, and the asterisk that flattered them. — 02:42

The only runnable model here is the The one nobody announced. — 03:44

Want to actually run it yourself? GLM 5.3 flash. And nothing else on this list. — 04:15

The meme named the model. The license decides who can touch it. — 04:47

Concepts and entities

Mistral Large 4, Kimi K3, GLM-5.3, GLM-5.3-Flash, Mixture of Experts, KDA attention, Moonshot, Z.ai, Mistral, Harvey, Artificial Analysis

Full transcript (original language)

00:00 Picture this, October, three tabs open in my browser. Mistral, Moonshot, Z.A.I. Three flagship launches in 12 weeks. Each labs chart shows its own model winning. My benchmark spreadsheet disagrees with itself in three fonts. I’ve lost a weekend to this. One of these APIs is cost 68 cents per million tokens. Another costs $3. Both claim victory. The fan spins up just thinking about the downloads. I assumed it was the usual YouTube thumbnail math, but it wasn’t. Meet the contenders. Mistral large four, and yes, they named it Le

00:33 Chonk. 1.05 trillion parameters, 49 billion active per token, launched October 6th. Promo price, 68 cents in $2.09 out per million tokens. Kimmy K3 from Moonshot. 2.8 trillion parameters. The agenda champ. GLM 5.3 from Z.A.I. Roughly 743 billion parameters, plus an MIT licensed sibling called Flash. Three charts, three winners. If we stopped here, you’d think just pick the biggest number, but that’s the wrong way to look at it. Here’s the mechanism nobody puts on the slide. All three are mixture of experts models. Total parameters are the

01:05 rent. Active parameters are the speed. Picture a warehouse. You store every box, but the forklift grabs a few per order. ML4 stores 1.05 trillion, grabs 49 billion per token. Gam stores around 743 billion, grabs about 40. That’s why you pay compute for the active slice, not the warehouse. You pay for the slice, you store the whole thing. Context works the same way. All three claim 1 million tokens. Evaluators measured ML4 closer to 512K,

01:37 half the claim. Only K3 engineered for it. Its KDA attention decodes long context 6.3 times faster than K2. So, where does each one actually win? Once you see the pattern, you can’t unsee it. Broad quality, GLM 5.3. The independent valves index reads 53.51 for GLM, 50.30 for K3, 48.05 for ML4. Artificial analysis agrees. 45 44 3H GLM is number one of 112 open weight models. Coding, statistically a tie. Both Chinese models sit around 69% on DPW’s leaderboard.

02:07 Mistral’s own chart shows 61.7 at genic work K3 and it’s not close. Number one on tool of one 91.2 on browse comp. Top of the front end arena at 1679 LO and it’s only flagship here that sees image and video input. GLM 5.3 is text only. Mistral’s niche legal agents. 15.83% on Harvey. Best of the three, economics. GLM again. Same 69% coding but $3.99 per DPW task versus K3’s $4.65. And on her tries, GLM’s pass at four hits 87.6 beating Claude Fable 5’s 84.1. I know

02:42 you’re thinking so GLM wins. Hold that thought. I want to be up front. Every one of these charts is lawyered. Mistral’s claim is best open weight model developed outside China. That qualifier is load bearing. It’s third of this trio on both independent and indexes. GLM’s plus 50% coding jump measured on z.ai’s own internal benchmark. And K3 kills the cheap Chinese model myth outright. $3 in 15 out. That’s Claude at Sonnet territory. Hacker News ran a thread titled Kimi K3 is not cheap. Simon Willison’s test

03:13 Pelican cost 25 cents. Even the cyber numbers are partly refusal artifacts. The US and UK safety institutes tested open weight K3. Zero of 41 exploit tasks executed. Nobody lied here. They just picked the baseline, the harness, and the asterisk that flattered them. Here’s the constraint nobody mentions. You’re thinking open weights, I’ll self host. Up front honesty, none of the three flagships is runnable. ML 4’s weights run about 2.1 terabytes in BF 16 and don’t exist yet. K3 is 1.56 terabytes in native 4-bit. Moonshot recommends 64

03:44 plus GPUs. GLM’s flagship needs an HNOO node. The only runnable model here is the The one nobody announced. GLM5.3 My T license. 320 billion total. 18 billion active. 93 to 120 gigabytes quantized. Fits a 128 gigabyte Mac. Full disclosure, it scored zero of five DPX Spark. A serving bug capped its contacts at 24,576 tokens. Monthly bills at 10 million tokens a day. Mistral, $289. GLM, $600. Kimi, $1,620.

04:15 So, the verdict, split by who you are. Coding and agentic workloads, Kimi K3. Accept the Sonnet tier bill. Best quality per dollar? GLM 5.3. The index scores and the task costs both say so. Cheapest flagship API or you need EU hosting. Mistral large for $289 a month is hard to argue with. Want to actually run it yourself? GLM 5.3 flash. And nothing else on this list. Here’s the bigger picture. Open weights at the trillion parameter scale means nothing

04:47 to a hobbyist. You can download the file. You can afford the machine. The meme named the model. The license decides who can touch it. If this saved you a weekend of scratchy pain, subscribe. I test so you don’t have to.