Kimi K3 Complete Guide: The 2.8T World's Largest Open Model and the 0.18B Tiny Version That Runs on Potato PCs

Q: Does Kimi K3 run on an ordinary PC?

**The 0.18B tiny version runs** (~4GB CPU). But practical conversation isn't possible. The 2.8T main model and GGUF version don't run yet.

Q: Where can I get the 0.18B version?

HuggingFace: `inference-optimization/Kimi-K3-0.18B` (MIT license).

"A 2.8-trillion-parameter giant AI" and "a tiny AI that runs in 700MB" actually share the same DNA.

In July 2026, Moonshot AI (maker of Kimi) released Kimi K3 — the world's first "3T-class (trillion-parameter) open-weight model."

At the same time, the community released a 0.18B (180M parameter) tiny version that runs on ordinary PCs. The compression ratio from 2.8T to 0.18B is about 15,000× — while preserving the same architecture.

Kimi-K3 HuggingFace model card

In this article, we thoroughly explain both the official giant model and the community tiny version, based on real HuggingFace data and X posts.

Here's what you'll learn:

Kimi K3 specs and strengths (with benchmark comparisons)
Why the "giant model" and "tiny model" share the same DNA
The technical mechanics of the 0.18B version and its actual running size
Current status and cautions for running with llama.cpp etc.
Which one fits you

What Is Kimi K3 (Official, 2.8T Version)?

Kimi K3 is an open-weight, natively multimodal agent model released by Moonshot AI. It's their flagship: it understands text, images, and video within the same model, with a 1M-token context window.

Basic Specs (from the official model card)

Item	Spec
Total parameters	2.8T (2,800B)
Active parameters	104B per token
Architecture	MoE (Mixture-of-Experts)
Layers	93
Experts	896 (2 shared)
Active experts per token	16
Attention	KDA (Kimi Delta Attention) + Gated MLA
Context	1,048,576 tokens (1M)
Vision encoder	MoonViT-V2 (401M)
Quantization	MXFP4 weights / MXFP8 activations (quantization-aware training)
Modalities	Text, image
License	Kimi K3 License (open-weight)

What Is the New "KDA" Attention?

The biggest technical feature is Kimi Delta Attention (KDA) and Attention Residuals (AttnRes).

KDA: a new attention design that computes differences (deltas). Handles long contexts efficiently
AttnRes: retains attention residuals (subtraction differences), preventing information degradation
Stable LatentMoE: stabilizes the mechanism activating 16 of 896 experts

This achieves about 2.5× the scaling efficiency of Kimi K2 (previous generation).

Benchmarks (official scores)

Benchmark	Kimi K3	Claude Fable 5	GPT-5.6 Sol	GLM-5.2
Terminal-Bench 2.1	88.3	88.0	88.8	82.7
DeepSWE	67.5	70.0	73.0	46.2
SWE-Marathon	42.0	35.0	39.0	13.0
BrowseComp	91.2	88.0	90.4	—
GPQA Diamond	93.5	92.6	94.1	91.2
FrontierSWE	81.2	86.6	71.3	67.3

At 88.3% on Terminal-Bench 2.1, it's nearly level with GPT-5.6 Sol (88.8%). Top on SWE-Marathon at 42.0%. Performance on par with the world's frontier models.

What Is the "0.18B" Tiny Version? (Community)

According to X user @0x0SojalSec:

"A tiny Kimi-K3 that runs locally even on potato PCs.

From 2.8T to 0.18B.

0.10B active.

Same architecture.

Same new attention design scaling.

Same DNA, just shrunk.

Compressed to 0.18B for testing.

Currently fits in 700MB."

In other words: not just a truncated giant model — a "research/verification miniature" that preserves the architecture (KDA, MoE, AttnRes) while cutting parameters to the extreme.

Tiny Version Technical Specs (HuggingFace: inference-optimization/Kimi-K3-0.18B)

Based on the official model card (moonshotai/Kimi-K3), reduced as follows:

Parameter	Original (2.8T)	Tiny (0.18B)
num_hidden_layers	93	4
hidden_size	7168	512
num_attention_heads	96	4
intermediate_size (dense)	33792	1024
moe_intermediate_size	3072	256
num_experts	896	8
num_experts_per_token	16	2
num_shared_experts	2	1
kv_lora_rank	512	64

Important: What "Same DNA" Means

It's not just smaller — it maintains all layer types (KDA linear attention, MoE FFN, MLA full attention).

The tiny version's 4-layer structure:

Layer 0: Dense FFN + KDA linear attention
Layer 1–2: MoE FFN + KDA linear attention
Layer 3: MoE FFN + MLA full attention

This lets you test the behavior of the new attention design (KDA) itself on ordinary hardware.

Actual Running Size

Parameters: 0.18B (~0.10B active)
File size: ~700MB (F32)
License: MIT (the derivative is MIT, not the original Kimi K3 License)
Verified: perplexity=2.12 (passes the target of 10.0 or below)
Generation test: "According to all known laws of aviation..." — normal text generation confirmed

Another Miniaturization Route: GGUF Version (IQ1_S, 1bit/2bit)

X user @GrEarl converted the 2.8T model to GGUF format for llama.cpp.

Source: 1,453.8 GiB (MXFP4)
Output: 527.48 GiB (94-split GGUF)
Effective bitrate: 1.63 bpw (sub-2bit)
Method: routing experts at IQ1_S (~1.56bpw), everything else at Q4_K

But there's an important caveat (see "Current Limitations" below).

Current Limitations (Being Honest)

⚠️ Tiny version (0.18B) use cases

Strictly for testing, development, and architecture verification
Don't expect practical conversation or coding (0.18B has almost no knowledge)
A technical demo to confirm "whether the Kimi K3 architecture runs on an ordinary PC"

⚠️ GGUF version (IQ1_S) doesn't run yet

As @GrEarl himself states:

Cannot be loaded by released llama.cpp
Kimi K3 support is only in an unmerged PR (ggml-org/llama.cpp#26185)
"Structural integrity verified, but actual loading and generation unverified"
2bit/1bit fixes and 1bit release in parallel

In other words, "practical operation of Kimi K3 on an ordinary PC" is not possible right now. Architecture verification (0.18B) is possible; the 2.8T GGUF is a technical proof-of-concept.

Required Hardware (by Size)

Version	Size	Approx. memory	Use
2.8T (official, MXFP4)	~1.45TB	Multiple GPUs (dozens of H100-class)	Research, production API
GGUF IQ1_S	~527GB	600GB+ VRAM (requires unofficial branch)	Technical verification (not yet run)
0.18B (tiny)	~700MB	CPU ~4GB	Architecture verification, learning

FAQ

Q1. Does Kimi K3 run on an ordinary PC?

The 0.18B tiny version runs (~4GB CPU). But practical conversation isn't possible. The 2.8T main model and GGUF version don't run yet.

Q2. Why is 2.8T and 0.18B called "same DNA"?

Because it's shrunk while maintaining the layer types (KDA linear attention, MoE, MLA) and the new attention design. The architecture's behavior itself is the same.

Q3. Where can I get the 0.18B version?

HuggingFace: inference-optimization/Kimi-K3-0.18B (MIT license).

Q4. When will the GGUF version run?

Once llama.cpp PR (#26185) merges into official builds. Timeline TBD.

Q5. Is it good at Japanese?

The official 2.8T is multilingual. The 0.18B version has too few parameters to be practical.

Q6. Which model should I try?

If you're interested in the architecture, try 0.18B locally. For practical use, use 2.8T via API (Together AI etc.).

Q7. What's the license difference?

The 2.8T main model uses the "Kimi K3 License" (open-weight). The 0.18B derivative is MIT. The GGUF version inherits the Kimi K3 License.

Q8. What does it cost?

0.18B is free and local. The 2.8T API follows the provider's pricing.

Summary — How to Think About Kimi K3

Conclusion: Kimi K3 is both a historic milestone — the world's first 3T-class open model — and the beginning of an ecosystem where anyone can try the architecture.

Official 2.8T: frontier-class performance, e.g. Terminal-Bench 88.3%
Tiny 0.18B: verify the KDA architecture on an ordinary PC in 700MB
GGUF version: technical proof-of-concept (not yet verified running)

Conversely, these people can wait:

Those who want "smart Kimi" locally right now (0.18B isn't practical)
Those who want to run it like a normal local LLM (Qwen, Llama, etc.) (GGUF not yet supported)

Kimi K3's true value lies in the next step: whether "giant model intelligence" can be brought down to a size anyone can touch. The 0.18B is the first step.

"You can actually load the new Kimi K3 architecture on normal hardware now."

That's the front line of local AI in 2026.

Information in this article is as of July 28, 2026. Benchmarks based on Moonshot AI's official announcement (Kimi-K3 model card). For the 0.18B version, see inference-optimization/Kimi-K3-0.18B; for the GGUF version, see GrEarl/Kimi-K3-GGUF-IQ1_S.