Moonshot AIKimiKimiLinearForCausalLM

Kimi Linear 48B A3B Instruct — Hardware Requirements & GPU Compatibility

ChatFunctions

Kimi Linear 48B A3B Instruct is a 49.1B-parameter open language model from Moonshot AI in the Kimi family. At Q4_K_M it needs about 30.28 GB of VRAM — see which GPUs and Macs can run it below.

49.1B params · MoE3.5B activeMITOct 2025190.7K downloads609 likes22.6K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Usable on a 24 GB GPU

Runs on 24 GB (IQ3_XXS · ~275 tok/s · 16K)Macs from 24 GB (IQ2_S · ~93 tok/s)

MIT

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Kimi Linear 48B A3B Instruct

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on RTX A6000 (48 GB) (14 GB headroom) · Q4_K_M

Generation speed: ~173 tok/s; generation speed
Cost per 1M output tokens: $0.53; per 1M output tokens

Deployon RunPodBest optionfrom $0.33/hr

Compare GPUs →

Run Kimi Linear 48B A3B Instruct

33.9 GB VRAM at 16K context · needs offload on the RTX 4090 (9.9 GB short)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve moonshotai/Kimi-Linear-48B-A3B-Instruct --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~108 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model mlx-community/Kimi-Linear-48B-A3B-Instruct-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model mlx-community/Kimi-Linear-48B-A3B-Instruct-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX A6000 (48 GB)CheapestCheapest14 GB free	~173 tok/s	$0.53	Deploy on RunPodfrom $0.33/hr (opens in new tab)	48 GB · 14 GB free · ~173 tok/s · $0.53 / 1M tokens
A40 (48 GB)14 GB free	~160 tok/s	$0.61	Deploy on RunPodfrom $0.35/hr (opens in new tab)	48 GB · 14 GB free · ~160 tok/s · $0.61 / 1M tokens
RTX A6000 (48 GB)14 GB free	~173 tok/s	$0.65	Deploy on Vast.aifrom $0.40/hr (opens in new tab)	48 GB · 14 GB free · ~173 tok/s · $0.65 / 1M tokens
A40 (48 GB)14 GB free	~160 tok/s	$0.93	Deploy on Vast.aifrom $0.54/hr (opens in new tab)	48 GB · 14 GB free · ~160 tok/s · $0.93 / 1M tokens
RTX 6000 Ada (48 GB)14 GB free	~203 tok/s	$0.88	Deploy on Vast.aifrom $0.64/hr (opens in new tab)	48 GB · 14 GB free · ~203 tok/s · $0.88 / 1M tokens
RTX 6000 Ada (48 GB)14 GB free	~203 tok/s	$1.01	Deploy on RunPodfrom $0.74/hr (opens in new tab)	48 GB · 14 GB free · ~203 tok/s · $1.01 / 1M tokens
L40S (48 GB)14 GB free	~188 tok/s	$1.17	Deploy on RunPodfrom $0.79/hr (opens in new tab)	48 GB · 14 GB free · ~188 tok/s · $1.17 / 1M tokens
L40S (48 GB)14 GB free	~188 tok/s	$1.18	Deploy on Vast.aifrom $0.80/hr (opens in new tab)	48 GB · 14 GB free · ~188 tok/s · $1.18 / 1M tokens

Compare GPUs →

How Much VRAM Does Kimi Linear 48B A3B Instruct Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
14.3 GB	22 GB	46.4 GB	12.1 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 32K context · ~330 tok/s
15.6 GB	23.2 GB	47.7 GB	13.9 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 32K context · ~316 tok/s
16.2 GB	23.8 GB	48.3 GB	14 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 32K context · ~309 tok/s
17.4 GB	25 GB	49.5 GB	15.8 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 16K context · ~297 tok/s
19.8 GB	27.5 GB	52 GB	19.7 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 16K context · ~275 tok/s
20.5 GB	28.1 GB	52.6 GB	16.7 GB	2-bit · small K-quant — 2-bit small K-quant	Fits: 8K context · ~270 tok/s
21.1 GB	28.7 GB	53.2 GB	20.4 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 8K context · ~266 tok/s
21.7 GB	29.3 GB	53.8 GB	17.5 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 8K context · ~261 tok/s
21.7 GB	29.3 GB	53.8 GB	21.3 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 8K context · ~261 tok/s
22.3 GB	29.9 GB	54.4 GB	21.6 GB	3-bit small — 3-bit small quantization	Fits: 8K context · ~257 tok/s
22.9 GB	30.6 GB	55 GB	22.7 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 4K context · ~252 tok/s
24.8 GB	32.4 GB	56.9 GB	22.7 GB	3-bit medium — 3-bit medium quantization	Does not fit: needs 28.3 GB
25.4 GB	33 GB	57.5 GB	28.4 GB	4-bit · legacy — 4-bit legacy quantization	Does not fit: needs 28.9 GB
26.0 GB	33.6 GB	58.1 GB	23.6 GB	3-bit large — 3-bit large quantization	Does not fit: needs 29.6 GB
27.2 GB	34.9 GB	59.3 GB	26.5 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Does not fit: needs 30.8 GB
28.4 GB	36.1 GB	60.6 GB	29 GB	4-bit small — 4-bit small quantization	Does not fit: needs 32 GB
28.4 GB	36.1 GB	60.6 GB	31 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Does not fit: needs 32 GB
28.4 GB	36.1 GB	60.6 GB	28 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Does not fit: needs 32 GB
30.3 GB	37.9 GB	62.4 GB	30.1 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Does not fit: needs 33.9 GB
30.9 GB	38.5 GB	63 GB	30.3 GB	4-bit large — 4-bit large quantization	Does not fit: needs 34.5 GB
34.6 GB	42.2 GB	66.7 GB	34 GB	5-bit small — 5-bit small quantization	Does not fit: needs 38.2 GB
35.8 GB	43.5 GB	67.9 GB	35.1 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 39.4 GB
36.4 GB	44.1 GB	68.5 GB	35.3 GB	5-bit large — 5-bit large quantization	Does not fit: needs 40 GB
41.3 GB	49 GB	73.5 GB	40.5 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 44.9 GB
49.9 GB	57.6 GB	82 GB	52.2 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 53.5 GB
99.0 GB	106.7 GB	131.2 GB	98.3 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 102.6 GB

Selected Q4_K_M

4-bit medium quantization

33.9GBat 16K context

Weights: 29.5 GB
KV cache: 4.1 GB
Overhead: 0.3 GB

Doesn't fit RTX 4090 (reference) · needs 9.9 GB more, IQ3_XXS fits

≥50 tok/s needs ≥178 GB/s · M3 Ultra ~109, M2 Ultra ~108

Or rent RTX A6000 from $0.33/hr.

Which Hardware Can Run Kimi Linear 48B A3B Instruct?

Q4_K_M · 30.3 GB

Kimi Linear 48B A3B Instruct (Q4_K_M) requires 30.3 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 40+ GB is recommended. 1 GPU can run it, including NVIDIA GeForce RTX 5090. 78 devices with unified memory can run Kimi Linear 48B A3B Instruct, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M5 Max (48 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8–16 GB · RTX 4060, RTX 3060 12GB, RTX 5060 Ti 16GBDoesn't fit — smallest quant needs 17.9 GB8 GB tier →12 GB tier →16 GB tier →

24 GB · RTX 4090referenceIQ3_XXS · 23.4 GB · ~275 tok/s · 16K23.4~27516K24 GB tier →

Mac 32 GB · Mac Mini M4Q3_K_M · 28.3 GB · ~38 tok/s · 16K28.3~3816KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 33.9 GB · ~91 tok/s · 32K33.9~9132KMacs →

Alternatives to Kimi Linear 48B A3B Instruct

Kimi family · 5 models →

Similar models

4tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Mixtral 8x7B Instruct v0.1Mistral AI	65.1	23.5 GB	~92	same size
Llama 3 3 Nemotron Super 49B V1 5NVIDIA	—	22.8 GB	~35	same size
Nemotron H 47B Reasoning 128KNVIDIA	—	107.2 GB	—needs > 24 GB	same size
AI21 Jamba Mini 1.5AI21 Labs	—	117.8 GB	—needs > 24 GB	same size

Fine-tunes & variants(1)community builds of Kimi Linear 48B A3B Instruct

Kimi Linear Tiny Random

Specifications

Publisher: Moonshot AI
Family: Kimi
Parameters: 49.1B
Active parameters: 3.5B per token
Architecture: KimiLinearForCausalLM
Vocabulary Size: 163,840
Release Date: 2025-10-30
License: MIT

Resources

Frequently Asked Questions

How much VRAM does Kimi Linear 48B A3B Instruct need?

Kimi Linear 48B A3B Instruct requires 30.3 GB of VRAM at Q4_K_M, or 99.0 GB at BF16.

VRAM = Weights + KV Cache + Overhead

Weights = 49.1B × 4.8 bits ÷ 8 = 29.5 GB

KV Cache + Overhead ≈ 0.8 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

VRAM usage by quantization

Q4_K_M

30.3 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run Kimi Linear 48B A3B Instruct?

Yes, at IQ3_M (22.9 GB) or lower. Higher quantizations like Q3_K_M (24.8 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for Kimi Linear 48B A3B Instruct?

For Kimi Linear 48B A3B Instruct, Q4_K_M (30.3 GB) offers the best balance of quality and VRAM usage. Q4_K_L (30.9 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 14.3 GB.

VRAM requirement by quantization

IQ2_XXS

14.3 GB~53%

IQ3_XS

21.1 GB~73%

Q3_K_L

26.0 GB~86%

Q4_K_M ★

30.3 GB~89%

Q4_K_L

30.9 GB~90%

BF16

99.0 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Kimi Linear 48B A3B Instruct on a Mac?

Yes — Mac Mini M5 Pro (24 GB) and 106 other Macs can run Kimi Linear 48B A3B Instruct, the smallest at IQ2_S. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 30.3 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M5 Max (36 GB).

Can I run Kimi Linear 48B A3B Instruct locally?

Yes — Kimi Linear 48B A3B Instruct can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Kimi Linear 48B A3B Instruct?

At Q4_K_M, Kimi Linear 48B A3B Instruct can reach ~171 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 2.1 GB active ÷ (8000 × 0.65) = 0.413 ms, plus 27 layers × 0.055 ms = 1.485 ms, so 1000 ÷ 1.898 ms = ~527 tok/s

Estimated speed at Q4_K_M (30.3 GB)

NVIDIA B200

~527 tok/s

NVIDIA B300

~527 tok/s

NVIDIA GH200 Grace Hopper Superchip

~463 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Kimi Linear 48B A3B Instruct?

At Q4_K_M, the download is about 29.47 GB. The full-precision BF16 version is 98.25 GB. The smallest option (IQ2_XXS) is 13.51 GB.

Which GPUs can run Kimi Linear 48B A3B Instruct?

1 consumer GPU can run Kimi Linear 48B A3B Instruct at Q4_K_M. Top options include NVIDIA GeForce RTX 5090.

Which devices can run Kimi Linear 48B A3B Instruct?

78 devices with unified memory can run Kimi Linear 48B A3B Instruct at Q4_K_M (30.3 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.