How Much VRAM Does Kimi K3.1 Pro Need?
As of October 31, 2026, the honest answer has two halves.
The weights do not exist yet. Moonshot announced Kimi K3.1 Pro on October 23 and the model runs today through the Kimi API, but the public line is "Full model weights will be released by Nov 4, 2026", and the Hugging Face repo for it still returns a 401. Any page telling you to download it before the files land is describing the API.
The size is already public. 2.8 trillion parameters, 104 billion active per token, a 1M-token context, text and images in. Moonshot calls it "a post-trained version of Kimi K3, with substantially larger post-training compute", and The Decoder reports the dimensions as unchanged from K3. So the resident math can come straight from K3's measured files: about 1,561 GB of weights in native MXFP4, about 1.8 TB once cache and runtime overhead are counted, and a million-token context that costs under 2% of the bill. The weights are the wall.
Locally there is nothing to run today: no files to download, no quant builds to load, and neither of the two expert-offload engines that put 552B-class models on desktops today lists a Kimi model.
Kimi K3.1 Pro VRAM Requirements at a Glance
These are K3's verified numbers, which K3.1 Pro is expected to inherit because it is a post-trained K3. The file list settles it, and this table gets updated the day it lands.
| Component | At 8K context | At the full 1M context |
|---|---|---|
| Weights, native MXFP4 (measured for K3) | 1,561 GB | 1,561 GB |
| KV cache | 0.7 GB | 29.5 GB |
| Runtime overhead (~15%) | ~234 GB | ~239 GB |
| Resident total | ~1,796 GB | ~1,829 GB |
Scale lines, so the number means something. Thirteen H200s at 141 GB each is 1,833 GB, and that pool only holds the model; Moonshot's own K3 launch post recommended "supernode configurations with 64 or more accelerators", and its stated reason is communication bandwidth between accelerators, not capacity. On 24 GB consumer cards you are counting to 75 before the overhead starts. And the column that surprises people is context: the full 1M window costs 29.5 GB on a model carrying 1,561 GB of weights. Long context is not what locks you out. The download is.
What Is Confirmed and What Is Not
Confirmed, from Moonshot's announcement and Artificial Analysis's launch measurement, both dated October 23:
| Spec | Value | Whose word |
|---|---|---|
| Total parameters | 2.8T | Vendor (X post), independent (AA) |
| Active per token | 104B | Vendor (X post), independent (AA) |
| Context | 1M tokens | Vendor, AA |
| Inputs | Text and image | Vendor, AA |
| What it is | A post-trained K3 with much larger post-training compute | Vendor blog |
| Open weights | Promised by Nov 4, 2026 | Vendor blog and X post |
| API price | $2.00 input, $10.00 output, $0.20 cache hits per million tokens | AA, Willison, The Decoder |
| Highspeed variant | 2 to 3 times faster, twice the price except cache hits | Vendor, Willison |
Moonshot's own copy says "3T-class". The number is 2.8T. It is the same rounding we untangled on K3 in July.
Not confirmed, and the article would be guessing if it filled these in:
| Open question | What we can say today |
|---|---|
| File size and format | Unknown. If the MXFP4 files ship again, expect near K3's measured 1.561 TB. FP8 arithmetic says about 2.8 TB, BF16 about 5.6 TB. The file list decides |
| The config | Layers, the MLA and KDA split, expert counts, cache dimensions: assumed to be K3's until config.json says otherwise |
| Licence | Unknown. K3 itself ships a custom Kimi K3 License, not MIT, with a Model-as-a-Service condition |
| Engine support | Neither the Strata calculator nor the Project Maya calculator lists a Kimi model; both pages read October 31, 2026 |
| Local speed and quality | Nothing to measure until the weights exist |
The Benchmarks: Two Tables, Two Owners
Artificial Analysis put K3.1 Pro 4th on its Intelligence Index at 55, 11 points above K3's 44 (max), behind only Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Muse Wonder Full (56). Those are Artificial Analysis's own runs, published October 23:
| Artificial Analysis | K3.1 Pro | Reference |
|---|---|---|
| Intelligence Index | 55 | 4th; K3 was 44 |
| Terminal-Bench 4.0 | 67% | Highest of any model on AA's run; K3 scored 13% on the same test |
| GDPval-AA | 1,846 Elo | 3rd, up 312 from K3; Opus 5.5 is 1,866 |
| AutomationBench-AA | 73% | Up from 58% |
| HLE | 52% | Up from 47% |
| CritPt | 25% | Up from 23% |
| SciCode | 61% | Up from 59% |
| AA-Omniscience | 30 | From 20; hallucination 53% to 44%, accuracy 48% to 51% |
| Cost per task | $1.98 | 55K output tokens per task, 14% more than K3's 48K |
Moonshot's launch table, the vendor's own figures:
| Vendor table | K3.1 Pro | Reference |
|---|---|---|
| Terminal-Bench 4.0 | 65.7% | GPT-6 Astra 58.2%, Claude Opus 5.5 66.4% |
| Kimi Code Bench 3.0 Long-horizon | 61.2% | K3 scored 55.95% |
| OfficeQA Pro v2 | 64.4 | Their headline knowledge-work row |
| Internal finance suite | +7.2 modeling, +11.6 research | Business operation about 4.2x K3 |
The Terminal-Bench gap is the interesting part, and it reconciles without anyone being wrong: AA's harness reads 67%, Moonshot's own run reads 65.7%. Two measurements by two parties. AA reports 67% as the highest score of any model on its run. Moonshot's own table puts 65.7% ahead of GPT-6 Astra's 58.2 and narrowly behind Claude Opus 5.5's 66.4. The harnesses differ, so we do not average them or line Moonshot's Opus figure up against AA's 67%. We publish both and attribute each.
The economics row matters for anyone budgeting API work: token prices came down 33% from K3, and cost per Intelligence Index task held at $1.98 because the model writes 14% more tokens per task. Simon Willison's line on the launch: "This is the first time a (soon-to-be) open weight model has scored this highly against the frontier." Artificial Analysis's headline says the same: it will lead open weights models once the weights are released.
What none of these tables contain: local throughput, quality at any quantization tier, or a single number from a machine you could buy at a computer shop. Those join on or after Nov 4.
The Resident Math
Weights. K3's published files measure 1.561 TB, which is 4.46 bits per parameter once the unquantized tensors are counted. A post-trained version of the same dimensions inherits that number. If K3.1 Pro ships the same MXFP4 format, scale nothing; if it ships wider, the first row of the table is the only line that moves.
The dropdown is not your rescue. The vendor already ships at 4-bit. K3's native files measure 1.561 TB at 4.46 effective bits per weight, and our calculator treats that as the floor: Q4_K_M's 4.85 and even Q8_0's 8.5 clamp down to the native figure, so the residency total stays near 1,796 GB whichever tier you pick. The real files agree. Unsloth's lossless Q8 build for K3 lands at 1.56 TB and its Q4 build at 1.51 TB, each within about 3% of the original. The ladder between 4 and 8 bits has nothing to sell you. The only exit is the 1 to 2 bit tier, and that is the next section.
Context is cheap. K3 keeps 69 of its 93 layers in Kimi Delta Attention, a fixed recurrent state that does not grow with context, so only 24 MLA layers pay per token: 0.7 GB at 8K, 29.5 GB at a full 1M. Under 2% of the bill. The K3 deep dive walks the split layer by layer, and the per-model breakdown runs the same math for every tier.
The 1-Bit Path, If the Quantizers Repeat K3's Ladder
There are no K3.1 Pro quantizations because there is no K3.1 Pro to quantize. What exists is Unsloth's K3 ladder from July, which shows a 2.8T model shrunk toward desk sizes. Its sizes were read from the Hugging Face repo on October 31, and our run-locally page covers what it does. If the same work lands for K3.1 Pro, this is the arithmetic that follows, and until then treat every "K3.1 Pro 1-bit" claim you see as describing K3's files or nothing.
| K3 build | Size | What it needs |
|---|---|---|
| UD-IQ1_S | 594 GB | About 683 GB pooled with overhead; 6 x 128 GB unified, or 4 B200s (768 GB) |
| UD-IQ2_XXS | 711 GB | 768 GB stops being enough once the ~15% overhead is counted (about 818 GB); plan 8 x 128 GB |
| UD-Q2_K_XL | 861 GB | About 990 GB with overhead; 8 x 128 GB (1,024 GB), or a small server |
| UD-Q4_K_XL | 1.51 TB | About 1,735 GB with overhead; 14 x 128 GB (1,792 GB), or 13 x H200 (1,833 GB) |
Unsloth reports about 78.9% top-1 agreement with the original for UD-IQ1_S, at 38% of its size, and its documentation says the 594 GB build "will require at least 610GB RAM". The Hugging Face repo also holds three smaller files, UD-Q1_0 at 466 GB, UD-TQ1_0 at 509 GB and UD-TQ2_0 at 552 GB, which the docs page does not size. All of this is K3. If the ladder repeats for K3.1 Pro, the entry ticket is a download in the 500 GB class.
The Offload Reality: Strata and Project Maya Today
This is the part that is real today. Strata and Project Maya run MoE models by keeping the hot experts in VRAM and RAM and reading the rest off the SSD per token. The lists, read from both tools on October 31, 2026: Strata 0.2.9 sizes Qwen 3.8 Flash Next, Qwen 3.6 35B-A3B and DeepSeek V4.1 Flash 1017; Maya 1.1.44 sizes DeepSeek V4.1 Flash 1017, GLM 5.3 Flash, MiniMax M3.1 Flash and Qwen-3.8-Flash-Next. No Kimi model on either list, and the release notes through Strata 0.2.9 and Maya 1.1.45 do not name one either.
What those lists have already proven, with named testers and their published numbers:
- @lyesbahi posted the 552B backbone of DeepSeek 1017, an ISTA 2.25 bpw build, on one RTX 5090 (32 GB) plus 121 GB of DDR5 on Strata 0.2.7 (engine 0.1.52): 47 tok/s on a 2,200-token generation, 33 tok/s at 256 tokens out. Long generations matter for agent work, and the long-generation row is the fast one.
- @pkillsleep posted MiniMax M3.1 Flash, a 428B model, from the 93 GB Maya-XS build on Maya 1.1.32 across an RTX 3090 plus an RTX 3080 Ti and 80 GB of RAM: 13.7 tok/s generation with 22% of the experts in VRAM, output length not stated. His own verdict was that the model is too slow for that setup, and he planned to retry with 128 GB of RAM.
Now the arithmetic on that same path, and it is arithmetic, not a measurement. The active set is 104B parameters; at a 1-bit class average of about 1.7 bits per weight, one token reads roughly 22 GB of expert weights. From an SSD at 5 GB/s that is about 0.23 tok/s. From system RAM at an optimistic 19 GB/s of effective random reads it is about 0.86 tok/s. A hot-expert cache moves the answer, but caching works when the hot set is small relative to the pool, and K3.1 Pro is 5 times the largest model these engines size today and 6.5 times the MiniMax that ran on those two cards. While the files stay unpublished, no measurement can outrun that arithmetic.
For it to get there anyway, four things have to happen: kernels for the 69 constant-state KDA layers, MXFP4 expert loading inside the engine, a quantizer ladder from someone like Unsloth, and a storage budget of 594 GB to 1.56 TB for the files alone. Until then, the nearest thing that actually runs in this weight class is K3 at 1-bit, and even that is a 594 GB download. The streaming rules behind all of it are on our weight-streaming reality check page.
What to Run Today
- Kimi K3 itself, if you have 610 GB or more of pooled memory and want the 2.8T experience at 1-bit: what it takes and what it does.
- DeepSeek 1017, MiniMax M3.1, GLM 5.3 Flash or Qwen 3.8 Flash Next on one desktop, offloaded through Strata or Maya: the Strata calculator and the Project Maya calculator size your exact cards.
- Rent the supernode, if the answer you need is a frontier 2.8T today rather than in a month: our GPU rental guide.
- Nothing at all for K3.1 Pro local. Moonshot's API and apps are the way to use it until the files are public. For what your own box does run, the GPU checker answers in seconds, and the Chinese open-weight map covers the whole field, not just Kimi.
What Changes on Nov 4
Moonshot's promise line is dated, so this page is too. On Nov 4, 2026, on this same URL, the placeholder updates in place:
- The repo gets read and the file list summed, replacing "expect about 1.56 TB" with the measured download.
- config.json replaces the inherited geometry: layers, the attention split, expert counts, cache dimensions, context.
- The licence gets named instead of flagged unknown.
- Quant builds get tracked as they ship, same ladder discipline as K3.
- Engine support gets re-checked on both tools before the text changes.
- Any local number we can stand behind arrives with its receipts: machine, backend, quant, context. Anything else stays attributed to whoever measured it.
- The calculator gets its K3.1 Pro entry, and we have held the K3 rule for this family: the entry goes live the day the config does.
If the weights slip, the promise line and this section get re-dated, and the page says so.
Sources
- Moonshot tech blog, "Introducing Kimi K3.1 Pro", 2026-10-23: https://www.kimi.ai/blog/kimi-k3-1-pro
- Moonshot's announcement post, 2026-10-23: https://x.com/Kimi_Moonshot/status/2113680701770068067
- Artificial Analysis launch benchmarks, 2026-10-23: https://artificialanalysis.ai/articles/kimi-k3-1-pro-launch-benchmarks
- Simon Willison on the launch, 2026-10-23: https://simonwillison.net/2026/Oct/23/kimi-k31-pro/
- The Decoder, 2026-10-23: https://the-decoder.com/chinese-ai-model-kimi-k3-1-pro-is-closing-in-on-anthropics-claude-opus-5-5
- Kimi K3 weights and config, the base model: https://huggingface.co/moonshotai/Kimi-K3
- The K3.1 Pro repo as of today (401, unpublished): https://huggingface.co/moonshotai/Kimi-K3.1-Pro
- Our own verified pages: https://vramcalculator.com/kimi-k3-vram-requirements/ and https://vramcalculator.com/run-kimi-k3-locally/
FAQ
Is Kimi K3.1 Pro open weights yet?
Not yet. Moonshot promises "Full model weights will be released by Nov 4, 2026", and as of October 31, 2026 the Hugging Face repository returns a 401. The model is live on the Kimi API and in Kimi's own products today. Any download guide published before the files exist is describing the API.
How much VRAM will Kimi K3.1 Pro need?
On K3's measured files, which K3.1 Pro is expected to inherit: about 1,796 GB resident at 8K context and about 1,829 GB at the full 1M window. That is 1,561 GB of weights, 0.7 to 29.5 GB of KV cache depending on context, and roughly 15% runtime overhead. The file list decides, and this page updates on Nov 4.
How big will the Kimi K3.1 Pro download be?
Unknown until the files land. If Moonshot ships K3's native MXFP4 again, expect about 1.56 TB, K3's measured figure. An FP8 build would be near 2.8 TB and a BF16 original near 5.6 TB by straight parameter arithmetic. The repo's file list settles it.
Can I run Kimi K3.1 Pro locally?
Today, no: the weights are not public. When they are, full residency is supernode-class hardware, and the 1-bit class build is how a model this size gets closest to a desk: for K3 that means a 594 GB download and at least 610 GB of combined memory. K3.1 Pro's builds depend on whether quantizers repeat that work after the files land.
Can Strata or Project Maya run K3.1 Pro with expert offloading?
Neither tool lists a Kimi model today; their lists cover Qwen, DeepSeek, GLM and MiniMax, read October 31, 2026. And the arithmetic is unforgiving: 104B active parameters at about 1.7 bits per weight means one token reads roughly 22 GB, about 0.23 tok/s from an SSD at 5 GB/s and about 0.86 tok/s from RAM at 19 GB/s. Caching helps, but this is 5 times the largest model those engines size today. Treat this as arithmetic, not a measurement.
What are the Kimi K3.1 Pro specs?
2.8 trillion total parameters, 104 billion active per token, Mixture-of-Experts, 1M-token context, text and image inputs. Announced October 23, 2026, live on the Kimi API, Kimi.ai, Kimi Work and Kimi Code, with a Highspeed variant at 2 to 3 times the speed.
Is Kimi K3.1 Pro just a post-trained K3?
That is Moonshot's own description: "a post-trained version of Kimi K3, with substantially larger post-training compute." The Decoder and Artificial Analysis both report the 2.8T and 104B dimensions as unchanged from K3. Until the config ships, "same geometry" is a named-source assumption, not a fact.
What does the Kimi K3.1 Pro API cost?
$2.00 per million input tokens, $10.00 output, $0.20 for cache hits, which Artificial Analysis puts 33% below K3's rates. The Highspeed variant doubles the price except on cache hits, which are priced the same. Cost per Intelligence Index task ran $1.98 on AA's measurement.
When will the VRAM calculator support Kimi K3.1 Pro?
The day its config.json does, the same rule we held when K3 launched. Until then the calculator sizes K3 and every other model in the catalogue, and the GPU checker tells you what your own hardware runs today.