Kimi K3.1 Pro
Kimi K3.1 Pro needs roughly 1749.1 GB VRAM with its native 4-bit checkpoint, the vendor's own release (1910.3 GB at a Q4_K_M estimate). 0 GPUs we track can run it fully in VRAM at 8k context.
0 GPUs run this natively · 0 with CPU offload
- NVFP4 total
- 1749.1 GB
- at 8k context
- KV cache, full context
- 51.5 GB
- 24 of 93 layers cache
- Inputs
- Text · Image
- Kimi K3.1 Pro, released October 2026
- Active experts
- 16 routed + 2 shared
- of 896 total, 104B active
- Real download
- 1.56 TB
- 96 MXFP4 safetensors shards
- Reasoning effort
- low · high · max
- max is the default
Kimi K3.1 Pro is a Mixture of Experts (MoE) model with 2800B total parameters but only 104B active per token developed by Moonshot AI. Launched through Moonshot's API on October 23, 2026, with the open weights following on November 4, Kimi K3.1 Pro is Kimi K3 with far more post-training: Moonshot describes large-scale joint reinforcement learning across hundreds of task domains. It keeps K3's size and shape: 2.8T total parameters with 104B active (16 of 896 experts plus 2 shared per token), 93 layers of which 69 are Kimi Delta Attention, a 401M-parameter MoonViT-V2 image encoder, and a native 1M-token window. The one architectural change is in the remaining 24 layers, which swap gated MLA for Gated DeepSeek Sparse Attention: each new token attends to the 2,048 earlier tokens a small indexer judges most relevant rather than to all of them, which makes long prompts cheaper to process. Reasoning effort is low, high, or max, and max is the default.
To run Kimi K3.1 Pro locally: this is a single-node model only if your node is eight 288 GB Blackwell GPUs. The native MXFP4 checkpoint is 1,561.3 GB of weights, and with KV cache and overhead it needs 1,749.1 GB at an 8k context. One B300 node (2,304 GB pooled) holds that with room for the full 1M-token window, which matches vLLM's own recipe of tensor-parallel 8 on one B300 node, one MI355X node, or two GB300 trays. Eight B200s (1,440 GB) or eight H200s (1,128 GB) are too small. The hybrid attention is why the window is nearly free once the weights fit: only 24 of 93 layers keep a per-token cache, so 1M tokens add 51.5 GB. No GGUF or local Ollama build existed a day after release, so there is no consumer path yet, and when one appears it will still be several hundred gigabytes. As a MoE model, inference speed depends on active parameters (104B) rather than total size.
On Moonshot's own table it trades blows with the closed frontier models at max effort: Terminal-Bench 4.0 65.7 against Claude Opus 5.5's 66.4 and GPT-6 Astra's 58.2, GPQA-Diamond 94.2, BrowseComp 92.1, OSWorld V2.1 79.6, and Humanity's Last Exam with tools 70.0. It is weaker on hard closed-book reasoning (HLE-Diamond 39.4, against 55.1 to 65.8 for those three closed models) and on cybersecurity. Against its own predecessor, the headline is reliability: on Moonshot's Hallu Bench Hard, the share of claims it gets wrong falls to 8.3%, from 23.4% for K3. All of these are Moonshot's numbers; the main independent measurement so far is Artificial Analysis, which scores it 46 on version 5.0.1 of its Intelligence Index and finds it slow and verbose at max effort, around 33 output tokens per second through Moonshot's own API.
What a million tokens of context costs on top of 1.56 TB
Kimi K3.1 Pro keeps K3's layer recipe: 69 of its 93 layers are Kimi Delta Attention, a linear-attention design whose memory is a fixed-size state that doesn't grow with the prompt. Only the other 24 store anything per token. The one change in K3.1 Pro is what those 24 layers do: Gated DeepSeek Sparse Attention picks the 2,048 most relevant earlier tokens for each new one instead of reading all of them, which saves compute on long prompts but still keeps every token's compressed entry in memory, so the cache is the same size as K3's.
KV cache only, at FP16. Weights and activation overhead sit on top of these figures.
Filling the whole 1,048,576-token window costs 51.5 GB of KV cache with the real hybrid stack, against 199.7 GB if all 93 layers cached. Stack that on the 1,561.3 GB native checkpoint and the full-context total is 1,806.4 GB, against 1,972.3 GB for the all-attention version, and only 57 GB more than the 1,749.1 GB the same weights need at 8k. In other words: on this model, going from a short chat to a million-token codebase barely changes the hardware bill. The weights decide everything.
VRAM at each quantization
Kimi K3.1 Pro natively supports a longer context window, but the table below is capped at 8k for comparability; its hybrid attention stack keeps KV cache growth well below linear.
| Quant | Weights | KV cache | Total |
|---|---|---|---|
| FP32 | 11200.0 GB | 0.40 GB | 12544.5 GB |
| BF16 | 5600.0 GB | 0.40 GB | 6272.4 GB |
| FP16 | 5600.0 GB | 0.40 GB | 6272.4 GB |
| Q8_0 | 2976.4 GB | 0.40 GB | 3334.0 GB |
| Q6_K | 2298.8 GB | 0.40 GB | 2575.1 GB |
| Q5_K_M | 1993.6 GB | 0.40 GB | 2233.3 GB |
| Q4_K_M | 1705.2 GB | 0.40 GB | 1910.3 GB |
| Q3_K_M | 1346.8 GB | 0.40 GB | 1508.9 GB |
| Q2_K | 1066.8 GB | 0.40 GB | 1195.3 GB |
| NVFP4reccuda | 1561.3 GB | 0.40 GB | 1749.1 GB |
KV cache figures assume 8k context at FP16. NVFP4 quantization requires a CUDA-capable GPU. Enable TurboQuant in the calculator to see reduced KV cache estimates.
Benchmarks
What the extra post-training bought over Kimi K3
Moonshot's launch post compares K3.1 Pro with K3 directly on only a few tests, mostly its own internal ones. These two are the ones it reports as absolute scores for both models; neither suite has a field in this site's benchmark schema.
Moonshot AI, Kimi K3.1 Pro launch post (kimi.ai/blog/kimi-k3-1-pro, 23 October 2026). Both are Moonshot's internal benchmarks, run at reasoning effort max; xdev-bench-hard is pass@3. No independent lab had reproduced either at time of writing.
Long-horizon coding moves about five points, from just under 56 (55.95) to 61.2. The exploit benchmark is the surprise: the released model scores 81.0 against K3's 33.3, even though Moonshot says it removed cyber-related data from both training stages for this open release, a step that lowered the score from 90.5 on the checkpoint trained without that filtering. The change Moonshot leads with is reliability rather than skill: on its own Hallu Bench Hard test, the share of claims K3.1 Pro gets wrong falls to 8.3%, from 23.4% for K3, which targets K3's best-known weakness: confidently wrong answers.
GPUs that run Kimi K3.1 Pro natively (0)
No single GPU in our list fits this model at Q4 with 8k context. Browse all GPUs or try multi-GPU setups.
Cloud GPU Rental
Cheapest way to run Kimi K3.1 Pro: 8× B300 288GB at NVFP4, ~24 tok/s, from $55.52/hr.
Pay by the hour · no contract · pods start in about a minute.
Rent this setup on RunPod ↗ (+$5 signup credit)Prices as of Oct 17, 2026. Rates change; check RunPod for current pricing.
Affiliate link: CanItRun may earn a commission. Doesn't affect the fit calculation above.
Notes
Moonshot's post-trained successor to Kimi K3, with much more reinforcement-learning compute and one architecture change: the 24 full-attention layers move from gated MLA to Gated DeepSeek Sparse Attention, which picks the 2,048 most relevant earlier tokens for each query instead of attending to all of them. The rest is K3's shape: 2.8T total parameters, 896 experts with 16 routed plus 2 shared per token (104B active), 93 layers of which 69 are Kimi Delta Attention, a 401M-parameter MoonViT-V2 vision encoder, and a native 1M-token window. Reasoning effort is low, high, or max (the default and the setting Moonshot benchmarked). The API launched October 23, 2026 and the weights opened November 4 under the custom Kimi K3.1 Pro License, MIT-style except that model-as-a-service businesses past $20M revenue over 12 months need a separate agreement, and products past 100M monthly users or $20M monthly revenue must show the model's name. The download is the MXFP4 (weights) / MXFP8 (activations) quantization-aware-trained checkpoint, about 1.56 TB; the higher-precision rows in the table are arithmetic, not files you can get. Moonshot says it removed cyber-related data from both training stages for the open release, which took its in-house exploit benchmark from 90.5 to 81.0. Benchmarks are Moonshot's own at max effort: GPQA-Diamond 94.2, Terminal-Bench 4.0 65.7 (Claude Opus 5.5: 66.4), BrowseComp 92.1, and a hallucination rate on its own hard test that falls from K3's 23.4% to 8.3%. Artificial Analysis independently scores it 46 on version 5.0.1 of its Intelligence Index, and measures it as slow and verbose at max effort: about 33 output tokens per second through Moonshot's API, with roughly two and a half minutes before the first answer token.
Compare Kimi K3.1 Pro with other models
How to run Kimi K3.1 Pro locally
NVFP4 needs 1749.1 GB: needs multiple datacenter-class GPUs (80 GB+ each).
vLLM 0.34.0+ (MXFP4, one 8-GPU B300 node)
vllm serve moonshotai/Kimi-K3.1-Pro \
--trust-remote-code \
--tensor-parallel-size 8 \
--load-format fastsafetensors \
--enable-auto-tool-choice \
--tool-call-parser kimi_k3 \
--reasoning-parser kimi_k3LM Studio: There is no GGUF build and no local Ollama tag yet: Ollama's kimi-k3.1-pro page offers only a cloud tag, so desktop frontends can only point at a remote endpoint. Moonshot recommends vLLM, SGLang, or TokenSpeed. vLLM's own recipe runs tensor-parallel 8 on one B300 node, one MI355X node, or two GB300 trays (add --nnodes 2, --node-rank, and --master-addr for the two-tray layout). For multi-turn chats and tool calls, send the complete assistant message back each turn, reasoning_content and tool_calls included, or the model loses its own reasoning.
Why this quantization? The 4-bit row is the model as Moonshot ships it, not a compromise. K3.1 Pro was quantization-aware trained to MXFP4 weights with MXFP8 activations, and the release is 96 safetensors shards totalling about 1,561.3 GB: the routed experts in MXFP4, everything else in BF16. The FP16, BF16, and Q8 rows above it are arithmetic, not checkpoints you can download, and the Q4 to Q2 rows are estimates for GGUF builds that don't exist yet.
Who is Kimi K3.1 Pro for?
Teams already running eight-GPU Blackwell-class nodes (B300, GB300, or MI355X) that want a frontier agentic model inside their own network, and anyone already serving Kimi K3, since the weights are the same size and the deployment carries over.
Best for
- Long-horizon coding agents working through a whole repository
- Office and finance agent work, where Moonshot reports its largest gains over K3
- Million-token context jobs on hardware that already holds the weights, since the full window adds only about 57 GB
- Self-hosted deployments where prompts and documents cannot leave your infrastructure
Not ideal for
- Any single GPU, workstation, or Mac, the weights alone are over 1.5 TB
- Hopper-generation nodes: eight H200s hold 1,128 GB, well short of the weights
- Latency-sensitive chat at max effort, where Artificial Analysis measured about two and a half minutes to the first answer token
- Teams that only need the capability, which the API sells at $2 per million input tokens
So, should you run it yourself?
Only if you already operate datacenter hardware, and then it is an easy upgrade. The smallest setup this site's calculator fits it on is one node of eight 288 GB B300s (2,304 GB pooled), which holds the native checkpoint with room for the full 1M-token window; eight 180 GB B200s (1,440 GB) fall short even at 8k. On that B300 node this site estimates about 24.2 tokens per second at short context. If you are already serving Kimi K3, the weights are the same size and the cache the same shape, so K3.1 Pro drops into the same deployment through vLLM or SGLang, with more reliable answers to show for it. Everyone else should use the API: Moonshot charges $2 per million input tokens and $10 per million output, and no consumer GPU, workstation, or 512 GB Mac Studio comes close to holding even a third of the weights.
Frequently asked questions
- What are the VRAM requirements for Kimi K3.1 Pro?
- Kimi K3.1 Pro's own release is a native 4-bit checkpoint that needs about 1749.1 GB of VRAM. On this site's standard quantization ladder it would need approximately 1910.3 GB of VRAM at Q4_K_M quantization, 3334.0 GB at Q8, and 6272.4 GB at FP16. These numbers assume 8k context window; its hybrid attention stack caches far fewer than all layers, so VRAM grows much slower than linearly with context.
- How many parameters does Kimi K3.1 Pro have?
- Kimi K3.1 Pro has 2800 billion total parameters, but only 104 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
- Can Kimi K3.1 Pro run on a 16 GB GPU?
- No. At Q4_K_M, Kimi K3.1 Pro needs 1910.3 GB of VRAM, more than 16 GB. You will need a multi-GPU server.
- Can Kimi K3.1 Pro run on a 24 GB GPU?
- No. Even at Q4_K_M, Kimi K3.1 Pro needs 1910.3 GB. Consider a multi-GPU server with 1911 GB+ of combined VRAM.
- What is the smallest quantization for Kimi K3.1 Pro that fits in 24 GB of VRAM?
- Kimi K3.1 Pro cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 1195.3 GB at Q2_K.
- What GPU do I need to run Kimi K3.1 Pro locally?
- You need a multi-GPU server. At Q4_K_M, Kimi K3.1 Pro needs 1910.3 GB VRAM, more than any single consumer GPU. That's roughly 24x 80 GB datacenter GPUs (H100, A100, or similar) pooled together.
- Can I run Kimi K3.1 Pro in the cloud?
- Yes. The cheapest RunPod setup we found that holds Kimi K3.1 Pro entirely in GPU memory at 8k context is 8× B300 288GB at NVFP4, at roughly 24.2 tokens per second and from $55.52 per hour for the whole pod. Prices as of Oct 17, 2026; RunPod rates change, so check their pricing page for the current figure.
- What hardware can actually run Kimi K3.1 Pro?
- About 1,749.1 GB at an 8k context with the native MXFP4 checkpoint, which is 1,561.3 GB of weights plus KV cache and runtime overhead. Filling the full 1,048,576-token window raises that to 1,806.4 GB. In practice that means one eight-GPU node of 288 GB B300s; a node of eight 180 GB B200s or eight 141 GB H200s is too small.
- Is there a GGUF or Ollama version of Kimi K3.1 Pro?
- Not a local one yet. Ollama's kimi-k3.1-pro page offers only a cloud tag, and a day after the weights opened on November 4, 2026, Unsloth's repository was a straight copy of the 1.56 TB safetensors with no GGUF files. For Kimi K3, Unsloth's GGUFs arrived two days after the weights and the recommended 1-bit build was about 594 GB, so even when K3.1 Pro builds appear they will need several hundred gigabytes of combined RAM and VRAM, and runtime support for the new sparse-attention layers first.
- What changed from Kimi K3 to Kimi K3.1 Pro?
- Mostly post-training: Moonshot says K3.1 Pro got substantially more post-training compute, including large-scale joint reinforcement learning across hundreds of task domains. The only architecture change is in the 24 full-attention layers, which move from gated MLA to Gated DeepSeek Sparse Attention for cheaper long-context compute. Parameter counts, experts, the 1M-token window, and the MXFP4 checkpoint format are unchanged, so the hardware needed to run it is the same as K3's.