Kimi K2 Instruct
Moonshot AI's 1-trillion-parameter mixture-of-experts (32B active per token). Trained on 15.5T tokens with a heavy emphasis on tool-use and agentic behaviour. Modified-MIT licence with an attribution clause for very-large deployments. Exceptional at long-horizon agent tasks; benchmarked well against Claude Sonnet on SWE-bench Verified.
- Parameters
- 1000B
- Context length
- 128K
- Modality
- text
- Released
- 2025-07-14
Memory & hardware
- VRAM (fp16)
- 2000 GB
- VRAM (Q4)
- 600 GB
- Recommended
- 8× H100 80GB at fp8, or hosted via Together / Groq
- Quantizations
- fp16, fp8, q4_k_m
License: Modified MIT (Kimi K2)
- SPDX
- —
- Commercial use
- Yes
- Modification
- Yes
- Redistribution
- Yes
Benchmarks
Hosted inference pricing
USD per million tokens.
| Provider | Input | Output | |
|---|---|---|---|
| togetherCheapest | $1.00 | $3.00 | |
| groq | $1.00 | $3.00 |
Run it yourself
Drop-in commands for the three most common open-source inference paths. The Ollama tag is a best-effort match against the registry; verify the size variant before pulling.
vllm serve moonshotai/Kimi-K2-Instruct
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("moonshotai/Kimi-K2-Instruct")
model = AutoModelForCausalLM.from_pretrained(
"moonshotai/Kimi-K2-Instruct", device_map="auto", torch_dtype="auto"
)moonshotai/Kimi-K2-Instruct Related models
Same family or similar size — useful when shopping around.
Reasoning model trained with reinforcement learning on top of DeepSeek V3-Base. MIT licence — even the weights are unrestricted, making R1 the most permissively-licensed frontier reasoning model. Generates long internal chains-of-thought before answering, trading latency for accuracy on math, code, and reasoning benchmarks. Distilled variants (e.g. R1 Distill Llama 70B) recover most of the quality at much smaller scales.
- Context
- 128K
- License
- mit
- VRAM Q4
- 402.6 GB
671B-parameter MoE model with 37B active per token. Trained for roughly $5.6M of compute — a landmark in cost-efficient frontier training. Frontier-class quality at a fraction of the cost of the closed proprietary frontier. The DeepSeek licence permits commercial use with limited restrictions on military and unlawful applications. Running V3 yourself requires serious hardware (8× H100 at fp8); most teams will use it via the DeepSeek API or providers like Together.
- Context
- 128K
- License
- deepseek
- VRAM Q4
- 402.6 GB
Meta's July 2024 flagship — the first open-weights model at 405B parameters. Trained on 15T tokens with 128K context. Rivals GPT-4o on many academic benchmarks and set the ceiling for open-weights quality for most of 2024. Running it self-hosted requires serious hardware (8× H100 at fp8 or multi-node at fp16); most users will run it via a hosted provider (Together, Groq, Fireworks). Llama 3.3 70B closed most of the practical gap at a fraction of the cost, so 405B is now most useful when 70B specifically hits its ceiling.
- Context
- 128K
- License
- llama-3
- VRAM Q4
- 243 GB
Hybrid Mamba-Transformer-MoE model with native 256K context (effective beyond 140K). 94B active parameters out of 398B total. The state-space-model layers give it linear-time scaling with sequence length, making it interesting for very long contexts. Licensed under AI21's open model licence, which permits most commercial use.
- Context
- 256K
- License
- jamba-open
- VRAM Q4
- 238.8 GB
NVIDIA's reward-modelling research vehicle. Trained primarily to be a synthetic-data-generation specialist rather than a chat-first model. Useful for teams building instruction-tuning datasets at scale.
- Context
- 4K
- License
- llama-3
- VRAM Q4
- 204 GB
xAI's first open-weights release: a 314B-parameter mixture-of-experts model. Apache 2.0 licensed. Largely a research artefact at this size — most users will run smaller models for production — but useful as a permissively-licensed reference for MoE research.
- Context
- 8K
- License
- apache-2-0
- VRAM Q4
- 188.4 GB