#5 Kimi K3
Strongest open weights, but not runnable on any home rig, so ranked for awareness, not practicality.
- Version
- Kimi-K3
- Maker
- Moonshot AI
- Release date
- 2026-06-13 (HF) (Hugging Face repo creation date unless noted)
- Architecture
- MoE (16 of 896 experts), Kimi Delta Attention, native vision
- Total params
- 2.8T
- Active params
- ~104B (Yotta Labs) unverified
- Context length
- 1M
- License
- Kimi K3 License (custom)
- Commercial notes
- Extra conditions only above $20M/yr revenue or 100M MAU (per Yotta Labs)
- Variants
- Kimi-K3 (MXFP4 QAT weights)
- Kimi K2.7-Code
- Kimi K2.6
- Kimi K3.1 Pro (reported 2026-10-23; weights not found on HF)
- Memory
- BF16/native: 1.56 TB MXFP4 (Yotta Labs)
FP8: no published figure yet
Q8: no published figure yet
Q4: 1.56 TB is already ~4-bit
Needs 1.6 TB+ aggregate GPU memory - Hardware
- Single Spark (128GB): No
Dual Spark (256GB): No
Quad Spark (~512GB): No
24GB GPU: No
128GB Mac / Strix: No
Exceeds even quad Spark - Tokens/sec
- no published figure yetSpeeds come from the cited sources. Entries marked "hosted, not local" are cloud API speeds, and "ESTIMATE" entries are estimates. Neither is a local measurement.
- Document Q&A strengths
- Best open model on BenchLeader (64.7)
- 1M context; BrowseComp 90.4 with full 1M ctx (model card)
- Leads coding/long agent loops
- Embedding / reranker pairing
- no published figure yet
- Runtimes
- Ollama: listed (kimi-k3; almost certainly cloud-only) unverified
llama.cpp: no published figure yet
vLLM: Yes (official recipe)
LM Studio: no published figure yet - Hugging Face
- https://huggingface.co/moonshotai/Kimi-K3
⚠️ The Ollama listing for this model may be a cloud-only tag, not local weights. Check before pulling.
DGX Spark configs it fits
None of the single, dual or quad Spark configs, based on published memory sizes.
Spark fit is estimated from published memory sizes, not tested on Spark hardware.
Sources
- https://huggingface.co/moonshotai/Kimi-K3
- https://www.yottalabs.ai/post/best-chinese-llm-models-2026-deepseek-qwen-glm-kimi-compared
- https://www.benchleader.com/best/open-weights-llm
- https://theopenweights.com/leaderboards/llm-intelligence
Last updated Oct 24, 2026. Fields tagged unverified lack a primary source. "No published figure yet" means we found no source, and we don't guess. Spark fit is estimated from published memory sizes, not tested on Spark hardware. Ollama listings may be cloud-only tags; check this for Kimi K3 and DeepSeek V4.1 before pulling.
The RAGShelf weekly
New local RAG tools, releases, and tested setups. One email a week, no spam.