Leaderboard / Moonshot AI (Kimi)

Kimi Linear 48B-A3B

Name: Kimi Linear 48B-A3B
Author: Moonshot AI (Kimi)

Moonshot AI (Kimi) · Kimi Linear · released 2025-10-30 · data checked 2026-11-01 Needs verification

Kimi Linear 48B-A3B is a 48B-parameter Mixture-of-Experts model (3B active per token) from Moonshot AI (Kimi) (China), released 2025-10-30 under the MIT license. It doesn't have enough reported benchmarks for an index score yet. It needs about 30.8 GB of memory at 4-bit with an 8K-token context.

Research model showcasing Kimi Delta Attention (hybrid linear attention).

Small LM Index

—

Needs 2 core benchmarks (has 0)

Parameters

48B

3B active · MoE

Price per 1M tokens

—

No hosted API on OpenRouter; self-host

Context window

tokens

Benchmarks

Each bar shows this model's score. The tick marks the best large model on that benchmark.

No benchmark results for this release yet. Know of one? Send it.

Specifications

Lab: Moonshot AI (Kimi) (CN · China)
Architecture: Mixture-of-Experts
Total parameters: 48B
Active parameters: 3B
Size class: Large
Context window: 1M tokens
Input modalities: text
Reasoning: Instruct (no thinking)
License: MIT
Openness: Open weights
Memory to run: ~30.8 GB at 4-bit, 8K context (more sizes)
Hugging Face: 76.5K downloads (30 days) · 613 likes

Weights on Hugging Face ↗

Compare with

Kimi Linear 48B-A3B vs Seed-OSS 36B Instruct (36.2B, index 60.3)
Kimi Linear 48B-A3B vs Hermes 4.3 36B (36.2B, index 51.4)
Kimi Linear 48B-A3B vs MiniCPM-V 4.7 35B-A3B (35.2B, index 54.5)

Sources

https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct

Something wrong here? Report it.

Run it locally

Uses bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF (Q4_K_M). Swap the tag for another size: BF16, IQ1_M, IQ1_S, IQ2_M, IQ2_S, IQ2_XS, IQ2_XXS, IQ3_M….

Ollama

ollama run hf.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q4_K_M

llama.cpp

llama-server -hf bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q4_K_M

Then open http://localhost:8080 for a chat window.

MLX (Mac)

pip install -U mlx-lm
mlx_lm.chat --model mlx-community/Kimi-Linear-48B-A3B-Instruct-4bit

vLLM (GPU server)

vllm serve moonshotai/Kimi-Linear-48B-A3B-Instruct

Full-precision weights, for a server GPU.

Open in LM Studio

Memory needed: weights, context cache and runtime buffers.

Quantization	4K context	32K context	256K context
16-bit F16	98 GB	98 GB	106 GB
8-bit Q8_0	53 GB	53 GB	61 GB
4-bit Q4_K_M	31 GB	32 GB	39 GB

Check it against your GPU, Mac or phone →