Kimi K3 in 8 GB of RAM: it's true, but count on 1.7 TB of disk and about an hour for a hundred words

In short

• The kimi-k3-in-c repo delivers on its memory promise: a measured peak of 8.24 GB of RAM, in C, with no graphics card and no math library. It had 11,215 stars on GitHub on 07/11 at 22:10 UTC.
• The price is paid in disk. The model weighs 1.56 TB and the repo asks for about 1.7 TB free. A 2 TB SSD cost between €319.95 and €399.95 on 07/11, depending on the model, at two online retailers.
• And in time. On the author's 124-core workstation, they measure 26.5 s per token with 8 GB of RAM, 5.6 s with 128 GB and up. For an answer of about 100 words, our calculation gives 66 minutes in the first case, 14 in the second.

The post vs. the repo

The post puts four pieces on the table. I went through them one by one, at the source.

The last row, the post doesn't give. The repo does. Right in its opening table.

A star, on GitHub, is a bookmark: a user marking the project with one click. The repo was created on 01/08/2026, and its code is under the Apache 2.0 license. The model weights aren't in it. They come from Moonshot AI, under its Kimi K3 license, which requires a separate agreement for any company selling access to the model whose revenue, subsidiaries included, exceeds $20 million over 12 months.

How 1.56 TB fits in 8 GB of RAM

Kimi K3 is a mixture of experts. Its layers, except the first, each contain 896 small sub-networks, the experts, and wake up only 16 of them per token. Out of 2.78T parameters, about 104 billion work at each step, or 3.7% of the total.

The repo's author drew four decisions from that about where the bytes live. The experts already arrive in 4-bit in the shipped file: 1.447 TB, 93% of the file. They stay on disk and are read on demand. Most of the rest, the trunk, weighs 108.81 GB and is used for every token. It is laid out one layer after another in a 109 GB file, then read layer by layer, as is, without compression.

The best part is elsewhere. Across twelve memory budgets, from 8 to 224 GB, the engine outputs exactly the same tokens. To guarantee that bit for bit, the author wrote their own math routines. A BLAS library picks the order of its additions by itself, and that order can change from one version or vendor to another. The whole engine weighs 176 KB. It's watchmaker's work, and every measurement is published, noise included.

The bill, line 1: the clock

A token is a piece of a word. How many for a hundred words? Google gives a rule for its Gemini models: 100 tokens make about 60 to 80 English words. A hundred words is therefore 125 to 167 tokens. The two English texts generated in the repo come out a bit lower, 51 words for 56 tokens. I count 150 tokens. It's an estimate.

I multiply by the author's table. They measured on a 124-core machine, with a 4.6 GB/s NVMe disk, memory capped at each step.

Our calculation, 150 tokens per answer.

The README demo says it in its own way, on another of the author's machines: 228 GB of RAM capped at 8 GB, and a slower disk, 3.2 GB/s. Given the sentence "The capital of France is", the engine answers "Paris", then keeps going. Eight tokens in 261.5 seconds, peak at 8.24 GB. Four minutes and twenty-one seconds for a city name.

These minutes don't include reading your question, paid at the first token: 98 seconds for a 17-token question, at 128 GB, on that same machine.

The repo also publishes a test on a tiny computer, a Jetson Orin Nano Super board in the 8 GB class, with the experts on an external spinning hard drive. A single token, "Paris", in 15 min 49 s on average over four runs.

The bill, line 2: the disk

The repo writes it at the top of its prerequisites: the real barrier is storage. The model file weighs 1.56 TB, in 96 shards. Add the rewritten trunk, 109 GB. Total required: about 1.7 TB free, preferably on a fast local disk.

The download is counted in hours rather than minutes, the author warns. Our calculation: at 1 Gbit/s sustained end to end, 1.56 TB takes 3 h 28 min.

Then comes the line no post shows. At 8 GB, none of the trunk stays in memory: its 108.81 GB goes back through the disk at every token, plus 25.83 GB of experts: about 135 GB per token, as the author puts it. For a hundred words, our calculation gives about 20 TB read. Nearly thirteen times the entire 1.56 TB file. According to the author, between 41 and 61% of a generation's time is spent waiting on the disk. The same program ran at 70.62 s per token on a network volume, versus 31.71 s on a local NVMe.

If your disk doesn't have 1.7 TB free, here's the club's entry fee:

Prices displayed on 07/11/2026 around 22:13 UTC, in stock.

I haven't verified that these two SSDs keep up with the throughput of the author's disk. The timing table depends on it.

What the repo doesn't claim

The author is the first to say so. They write it in the conclusion: "The point was never the speed." At 8 GB it takes about half a minute per token, and pretending otherwise would be silly.

The author also lists what the project doesn't do:

• no measurement of answer quality: the figures measure the engine, not the model;
• no images: the vision part of Kimi K3 isn't coded;
• a third of its timing measurements is noise: three identical runs gave from 14.78 to 20.14 s per token.

And the diagnostic tool only offers the "laptop" setting from 10 GB of free memory. Below that remains an "ultra" setting, which the author presents as a proof of life, not as a setting for chatting.

For anyone who just wants answers, Moonshot AI also offers Kimi K3 through an API. White Rose has detailed the Kimi K3.1 Pro API bill and picked apart the Kimi K3.1 Pro weights published on Hugging Face.

The verdict

I have the receipt. The post read the first line: 8.24 GB. It's right.

The following ones, the repo prints itself. 1.56 TB of file. About 1.7 TB of free disk, or at least €319.95 of SSD in our two price checks if yours is full. 66 minutes for a hundred words at 8 GB, on a 124-core machine, by our calculation.

Verdict: an engineering feat, honest about its limits. A very poor chat machine. The RAM is sorted. What's left is buying time.

What we don't know yet

• Speed on a real laptop. The 8 to 64 GB steps were measured on a big machine whose memory was capped, with a 4.6 GB/s disk. The repo publishes no timing taken on a laptop; its only small machine, the Jetson board, read its experts from a spinning hard drive.
• Quality. No answer-quality test in the repo.
• The real figure in French. Our rule of 150 tokens for a hundred words is an estimate; Kimi K3's tokenizer wasn't measured on French.

Key terms

• Parameter: one of the numbers tuned during a model's training. Kimi K3 has 2.78T, or 2,780 billion.
• Token: a piece of a word, the unit a model reads and writes.
• RAM: the computer's working memory, fast and limited.
• NVMe SSD: a flash-memory drive with no moving parts; NVMe refers to the fast format of these drives.
• Mixture of experts: a model split into many sub-networks, of which only a small part works for each token.
• Trunk: in this repo, the part of the model used for every token, experts excluded.
• 4-bit: a compact way of storing each weight, about four times lighter than the usual 16 bits.
• BLAS: a family of ready-made math libraries for matrix multiplication.
• Capped memory: a limit imposed by the system, which prevents the program from using more than the announced budget.

About the source

• Repo: FareedKhan-dev/kimi-k3-in-c, created on 01/08/2026, Apache 2.0 license.
• Model: moonshotai/Kimi-K3, Moonshot AI, Kimi K3 license.
• Read: 07/11/2026, between 22:10 and 22:21 UTC (SSD prices rechecked at 22:21 UTC, unchanged).

Source passages

Pages read on 07/11/2026 (UTC). The README is the main branch version as of that date; the measurement data is in the repo's docs/data folder.

Show the 17 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4	"description": "A 2.78-trillion-parameter Kimi K3 running inference on a single CPU in 8.24 GB of RAM. Portable C99: no BLAS, no framework, no GPU." ; "stargazers_count": 11215 ; "created_at": "2026-08-01T09:29:38Z" ; "spdx_id": "Apache-2.0"	https://api.github.com/repos/FareedKhan-dev/kimi-k3-in-c
S2↑1↑2↑3↑4↑5↑6+15	"2.78T parameters" ; "1.56 TB checkpoint on disk" ; "8.24 GB peak RSS, measured" ; "176 KB the whole engine" ; "Kimi K3 inference in portable C99. No BLAS. No framework. No GPU." ; "8 tokens in 261.5 s, 32.69 s/token average / PEAK RSS for the whole run: 8.24 GB" ; "only 16 of its 896 experts per layer fire for any given token" ; "About 104 billion parameters are active for any given token, out of 2.78 trillion, which is 3.7 percent." ; "routed expert set : 82,432 x 17,547,264 = 1.447 TB" ; "which is 93 percent of the entire checkpoint" ; "108.81 GB is the per-layer dense trunk" ; "5,560 GB […] 1,560 GB […] 113.49 GB […] 8.24 GB" ; "675× reduction" ; "Every rung produced byte-identical output. Twelve budgets" ; "A BLAS library chooses its own blocking and reduction order, which can differ between versions, between vendors" ; "The gate is storage: the checkpoint is 1.56 TB." ; "~1.7 TB free / 1.56 TB checkpoint + 109 GB packed trunk, ideally on fast local disk" ; "1.56 TB, so hours rather than minutes." ; "shards : 96" ; "step 0 took 97.98 seconds and read 200.67 GB" (prompt of 17 tokens, --trunk-gb 110 --cache-gb 13) ; "The point was never the speed. At 8 GB it takes about half a minute per token, and pretending otherwise would be silly." ; "This repository contains no model weights" ; "Every measurement here comes from one workstation: a two-socket AMD EPYC 7763 with 124 cores and no SMT, 228 GB of RAM" ; "O_DIRECT cold : 3.2 GB/s" ; "The trunk is 108.81 GB re-read in full on every token. The experts are 25.83 GB per token." ; R4: "Kelsey and I am a certified teacher. I have experience tutoring after school at the middle school level. I have taught" (24 tokens) ; "routing each token through a small subset of its total parameters, which keeps inference fast and cheap relative to its size. The model is trained on a large corpus of" (32 tokens)	https://github.com/FareedKhan-dev/kimi-k3-in-c (raw README: https://raw.githubusercontent.com/FareedKhan-dev/kimi-k3-in-c/main/README.md)
S3↑1↑2↑3↑4↑5↑6	"Workstation: 124 cores […] 720 GB RAM, 4.6 GB/s local NVMe. Note the disk is faster than the published box's 3.2 GB/s" ; "Laptop-class rows are enforced with a cgroup MemoryMax cap on the workstation NVMe" ; "8 GB (laptop preset) / 26.5 / 8.73 GB" ; "32 GB / 24.2 / 19.5 GB" ; "64 GB / 19.8 / 51.6 GB" ; "128+ GB (trunk resident) / 5.59 / 179 GB" ; "the first token of any run also pays a one-time load and is higher"	https://github.com/FareedKhan-dev/kimi-k3-in-c/blob/main/docs/data/speed-2026-08.md
S4↑1↑2↑3↑4	"the trunk is re-read in full every token (108.81 GB) while only ~25.8 GB of experts are touched" ; "run 1 14.7760 / 2 14.6655 / 3 20.1408" ; "A third of the measurement is noise." ; "These are engine measurements […] They are not quality benchmarks." ; memory-ladder.tsv: "8 0 0.0 0.49 28 32.69 […] 25.83 8.24" (pin_layers 0, trunk_hit 0.0)	https://github.com/FareedKhan-dev/kimi-k3-in-c/tree/main/docs/data
S5↑	"Between 40.9 and 60.6 percent of wall clock is spent waiting on the disk, at every rung." ; "A run at 31.71 s/token on local NVMe and a run at 70.62 on a network volume differ by 2.2 times"	https://github.com/FareedKhan-dev/kimi-k3-in-c
S6↑	"elif [ "$AVAIL_GB" -ge 10 ]; then PRESET=laptop; EXPECT="~27 s/token"" ; "The ultra preset […] is a proof-of-life path for 8 GB-class machines, not an interactive-speed preset"	https://github.com/FareedKhan-dev/kimi-k3-in-c
S7↑1↑2↑3	"No vision. MoonViT-V2 is fully specified in config.json at 27 layers, and has zero code here." ; "No quality benchmark. No perplexity, no task eval."	https://github.com/FareedKhan-dev/kimi-k3-in-c
S8↑1↑2↑3	"It is a 2.8T-parameter model" ; "You can access Kimi K3's API on https://platform.kimi.ai by selecting kimi-k3" ; "Activated Parameters 104B" ; "MXFP4 weights / MXFP8 activations (quantization-aware training)" ; API: "total": 2779931837184	https://huggingface.co/moonshotai/Kimi-K3
S9↑1↑2	"If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars […] over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI"	https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE
S10↑	"For Gemini models, a token is equivalent to about 4 characters. 100 tokens is equal to about 60-80 English words."	https://ai.google.dev/gemini-api/docs/tokens
S11↑1↑2	LDLC, « Kingston SSD NV3 2 To »: "price": "319.95", "priceCurrency": "EUR", "itemCondition": "NewCondition", "availability": "InStock" ; « Samsung SSD 990 EVO Plus M.2 PCIe NVMe 2 To »: "price": "399.95" (same listing)	https://www.ldlc.com/fiche/PB00630082.html ; https://www.ldlc.com/fiche/PB00653108.html
S12↑1↑2	Materiel.net, « Kingston NV3 - 2 To »: "price": "319.95", "priceCurrency": "EUR", "itemCondition": "NewCondition", "availability": "InStock" ; « Samsung SSD 990 EVO Plus - 2 To »: "price": "399.95" (same listing)	https://www.materiel.net/produit/202407270003.html ; https://www.materiel.net/produit/202409160083.html
S13↑1↑2↑3	TUNING: "The engine moves ~135 GB per token at low budgets." ; Jetson: "one-token inference with the complete official Kimi K3 text checkpoint on an 8 GB-class Jetson" ; "it is not an interactive-performance claim" ; "Mean inference time was 949.4358 seconds (15m 49.4s)" ; "official checkpoint and routed experts on […] a rotating external HDD" ; CHANGELOG 1.1.0: "Verified with four full runs on a Jetson Orin Nano Super against the released checkpoint"	https://github.com/FareedKhan-dev/kimi-k3-in-c/blob/main/docs/TUNING.md ; https://github.com/FareedKhan-dev/kimi-k3-in-c/tree/main/docs/results/jetson-orin-nano-super ; https://github.com/FareedKhan-dev/kimi-k3-in-c/blob/main/CHANGELOG.md
R1↑1↑2↑3↑4	Our calculation: 100 words = 150 tokens (estimate, between 125 and 167 per S10, 110 per R4). 8 GB: 150 × 26.5 = 3,975 s = 66 min 15 s; ×100 = 110 h 25 min (4.6 days). 32 GB: 150 × 24.2 = 3,630 s = 60 min 30 s; ×100 = 100 h 50 min. 64 GB: 150 × 19.8 = 2,970 s = 49 min 30 s; ×100 = 82 h 30 min. 128 GB and up: 150 × 5.59 = 838.5 s = 13 min 58 s; ×100 = 23 h 17 min 30 s	— (in-house calculation)
R2↑	Our calculation: 1,560,936,091,448 bytes × 8 / 10⁹ bits per second = 12,487 s = 3 h 28 min, at a constant 1 Gbit/s	— (in-house calculation, based on S2)
R3↑	Our calculation: at 8 GB (0 trunk layers kept in memory), (108.81 + 25.83) GB = 134.64 GB read per token (the author: "~135 GB per token at low budgets" S13) ; × 150 tokens = 20,196 GB, about 20 TB ; 20.2 / 1.56 = 12.9	— (in-house calculation, based on S2S4S13)
R4↑	Our count: 21 words for 24 tokens and 30 words for 32 tokens (recounted during fact-checking), i.e. 51 words for 56 tokens, about 110 tokens for 100 words (two English texts generated in the README)	— (in-house calculation, based on S2)