{
  "data": {
    "author": "Vlad Cealicu",
    "faq": null,
    "kicker": "Blog · Run it locally · 25 October 2026",
    "lede": "Moonshot doesn't pitch Kimi K3 for your desk. It recommends 64 or more accelerators for serving it well, and the smallest setups shown running it are eight-GPU servers. Here's what it would take at home anyway, and what that costs.",
    "readingMinutes": 5,
    "words": 1094
  },
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/blog/run-kimi-k3-locally",
    "json": "https://www.anchorterminal.com/blog/run-kimi-k3-locally.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/blog/run-kimi-k3-locally.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/blog/run-kimi-k3-locally.min.md"
  },
  "markdown": "By Vlad Cealicu, 25 October 2026. 1,094 words, about 5 min to read.\n\nWhen Meta opened the weights of Muse Spark Open, its announcement said two Mac Studios were \"the best way to run a powerful model on your own desktop\", and I wrote that [the vast majority of us can't](/blog/run-muse-spark-open-locally). Moonshot did the opposite with Kimi K3. It recommends \"supernode configurations with 64 or more accelerators\" for serving it efficiently, and says nothing about desks [[1]](#ref-1). I appreciate that.\n\nThis is the second post in a series on what it takes to run an open model yourself. Of the very large open models this summer, Kimi K3 is the one people talked about most, with 11,810 likes and 5.7 million downloads on Hugging Face, about 922,000 of them in the last 30 days [[2]](#ref-2).\n\n## How big is Kimi K3?\n\n2.8 trillion parameters, with 104 billion active for each token. It's a mixture of experts with 896 experts, 16 picked per token, a context of a million tokens, and images as well as text [[2]](#ref-2). The API launched on 16 July and the weights followed on 27 July [[1]](#ref-1) [[2]](#ref-2). Moonshot released Kimi K3.1 Pro on its API on 23 October, The Next Web reports, and says the full weights will follow by 4 November [[16]](#ref-16) [[17]](#ref-17).\n\nThe part that matters at home is that Moonshot trained it to run in 4-bit from the fine-tuning stage on, so the official weights are already 4-bit, and they're still 1,561 GB [[2]](#ref-2). With Muse you could drop to 4-bit and halve the size. With K3 that discount is already taken, and Unsloth's 8-bit file is lossless at 1,561 GB because of it [[3]](#ref-3).\n\n- 1,561 GB: the official weights, already 4-bit\n- 594 GB: Unsloth's recommended 1-bit file, about 610 GB of memory to run\n- 466 GB: the smallest file anyone has published, quality unknown\n\nKimi K3's weights, 24 October 2026. Sources [2] [3].\n\n## What do the smaller files cost you?\n\nUnsloth publishes GGUF files from 466 GB up. They recommend the 594 GB 1-bit one, which agrees with the 8-bit model's top token about 78.9 per cent of the time, against 90.4 per cent for their 861 GB 2-bit file [[3]](#ref-3). The 466 GB file has no quality numbers published. There's also a 181 GB build that keeps 160 of the 896 experts, and its own card says it's \"a code-capable model with degraded factual recall. It is not a general-purpose K3\" [[4]](#ref-4).\n\nSize in GB\n\n| File | GB |\n| --- | --- |\n| 1-bit, smallest | 466 |\n| 1-bit, recommended | 594 |\n| 2-bit | 861 |\n| 4-bit | 1509 |\n| 8-bit, lossless | 1561 |\n\nEach series has a zero-based scale.\n\n## What hardware would hold it?\n\nThese are my sums, using about 120 GiB of usable GPU memory on a 128GB machine as in the Muse post, which is about 129 GB [[5]](#ref-5). The 256GB Mac Studio is $9,499 or $10,799, because Apple's own pages disagree on whether 256GB needs the bigger chip [[6]](#ref-6).\n\nFor the full 1,561 GB, seven 256GB Mac Studios would hold it, about $66,500 to $75,600 [[6]](#ref-6). Or thirteen 128GB Strix Halo mini PCs or DGX Sparks, or a CPU server with 2TB of RAM, which is about $10,000 in the cheapest, slower DDR4-2400 to $19,000 at the median price today [[7]](#ref-7). Two linked DGX Stations, at over $90,000 each, come to about 1.5 TB and fall just short [[8]](#ref-8).\n\nFor Unsloth's 594 GB 1-bit file, it's three 256GB Mac Studios at about $28,500 to $32,400, five Minisforum MS-S1 Max at about $19,600, or five DGX Sparks at about $34,750 [[5]](#ref-5) [[6]](#ref-6) [[9]](#ref-9) [[10]](#ref-10). Or one DGX Station, listed from $94,930 in August, which Unsloth says can run it, with 748 GB of memory between its GPU and CPU [[3]](#ref-3) [[8]](#ref-8). When I checked on 24 October, the 512GB Mac Studio, which would halve the Mac count, still wasn't on sale and had no price [[6]](#ref-6).\n\nWhat holding Kimi K3 costs, US dollars, lower prices\n\n| Setup | US dollars |\n| --- | --- |\n| One DGX Station, 1-bit | 94930 |\n| Seven 256GB Mac Studios, full weights | 66500 |\n| Five DGX Sparks, 1-bit | 34750 |\n| Three 256GB Mac Studios, 1-bit | 28500 |\n| Five MS-S1 Max, 1-bit | 19600 |\n| 2TB of DDR4 for a CPU server, the RAM alone | 10000 |\n\nEach series has a zero-based scale.\n\n## How fast would it be?\n\nNobody has published a good number on hardware a person owns. The fastest is about one token a second, on a 128GB MacBook Pro reading the experts it needs from four SSDs as it goes, posted by a company building that engine [[11]](#ref-11). ComputingForGeeks got 0.10 tokens a second on a rented server with four A100s and 2TB of RAM, with the GPUs between 0 and 1 per cent busy and the memory split in a way that slowed it down, so I'd read that as a broken setup rather than a limit [[12]](#ref-12). Nobody has published a measurement for a cluster of Mac Studios yet.\n\nIn theory a CPU server could do better. Each token reads about 58 GB of the full weights, so eight channels of DDR4-3200 top out at about 3.5 tokens a second and twelve channels of DDR5 at about 8. On the 1-bit file it's about 22 GB a token, so 9 to 21 [[5]](#ref-5). Nobody has shown that in practice yet.\n\n- 1: token a second, a MacBook Pro streaming from four SSDs\n- 0.1: token a second, a rented server with four A100s\n- 122: tokens a second across 16 users, eight B300s\n\nKimi K3's published speeds. Sources [11] [12] [13].\n\nThe 122 is imec's, on an eight-GPU B300 server running 16 coding agents at once, about 7.6 tokens a second each, and the same post says K3 doesn't fit on an eight-GPU B200 server [[13]](#ref-13). NVIDIA validated its own 1,610 GB build on eight B300s too [[14]](#ref-14). That's the smallest serving setup anyone has shown.\n\n## Is it cheaper to use the API?\n\nBy a long way. Moonshot charges $3 per million tokens in and $15 per million out, and the cheapest provider on OpenRouter charges $5.50 per million out [[15]](#ref-15). Seven Mac Studios at $66,500 buy about 4.4 billion output tokens from Moonshot, before you've paid for the electricity [[6]](#ref-6) [[15]](#ref-15).\n\n## What I'd do\n\nUse the API, and say thank you to Moonshot for publishing the weights anyway, because they let other providers compete on price and let researchers look inside. The Kimi K3 Licence is based on MIT with two conditions. A company with a model-as-a-service business and more than $20 million in revenue needs a separate agreement before commercial use, and products with over 100 million monthly users or $20 million a month in revenue have to show \"Kimi K3\". Internal use is exempt [[2]](#ref-2).\n\nIf your data can't leave the building, K3 is a server job, not a desk job. The next post in the series is GLM-5.3-Flash, which one big Mac can run.\n\nKimi K3 now has [its own listing](/tools/kimi-k3) graded by our open-weight checklist, with its builds, the memory each needs and the command that starts it, and [a page comparing it with GLM-5.3-Flash](/compare/glm-5-3-flash-vs-kimi-k3). [Your cousins](/blog/your-cousins-open-models-for-agents) is about which open models an agent should pick.\n\n---\n\nMore in [Run it locally](https://www.anchorterminal.com/blog/run-it-locally/index.md) · Previous model [Can you run Muse Spark Open locally? Hardware, RAM and cost](https://www.anchorterminal.com/blog/run-muse-spark-open-locally.md)\n\n- Older: [The Log, issue 3: found, leaked, let in](https://www.anchorterminal.com/blog/the-log/issue-3-found-leaked-let-in.md)\n- All posts: https://www.anchorterminal.com/blog/index.md\n\n## References\n\nRead on 24 October 2026.\n\n[1] Moonshot AI, Kimi K3, on the Kimi blog. Source for the API being available that day, the weights by 27 July and the recommendation of \"supernode configurations with 64 or more accelerators\" for inference efficiency. https://www.kimi.com/blog/kimi-k3\n\n[2] Moonshot AI, Kimi-K3 on Hugging Face, with its config and licence. Source for 2.8T total and 104B active parameters, 896 experts with 16 selected, the million-token context, quantisation-aware training from the SFT stage onward, the 1,561.0 GB repository, the weights on 27 July, the licence's two conditions, and the likes and the all-time and 30-day downloads. https://huggingface.co/moonshotai/Kimi-K3\n\n[3] Unsloth, Kimi K3 guide and unsloth/Kimi-K3-GGUF. Source for the file sizes, the memory each needs, top-1 agreement with the 8-bit model of 78.9 and 90.4 per cent, the 8-bit file being lossless and K3 running on a DGX Station. https://unsloth.ai/docs/models/kimi-k3\n\n[4] mlx-community, Kimi-K3-mlx-reap160-2bit. Source for the 181 GB pruned build keeping 160 of 896 experts and its card's warning. https://huggingface.co/mlx-community/Kimi-K3-mlx-reap160-2bit\n\n[5] Anchor Terminal, can you run Muse Spark Open locally, 15 October 2026. Source for counting about 120 GiB of usable memory on a 128GB machine. The machine counts, totals and bandwidth ceilings here are our arithmetic, from the sizes in [2] and [3], the prices in [6] to [10], 204.8 GB/s for eight channels of DDR4-3200 and 460.8 GB/s for twelve channels of DDR5-4800. https://www.anchorterminal.com/blog/run-muse-spark-open-locally\n\n[6] Daring Fireball, configurations and pricing, 25 August 2026, and Apple's Mac Studio specs and store pages. Source for $5,499 plus $4,000 for 256GB, Apple's specs page tying 256GB to the 36-core chip, which would make it $10,799, and the 512GB option \"coming late October\" with no price. https://daringfireball.net/2026/08/configurations_and_pricing_for_new_mac_minis_and_mac_studios\n\n[7] DatacenterDisk, server RAM prices, updated 23 October 2026. Source for DDR4 at $4.92 a GB for the cheapest, DDR4-2400 sticks, and $9.06 a GB at the median. https://datacenterdisk.com/server-ram/ddr4\n\n[8] Tom's Hardware, 23 August 2026, and buyzero.de, read 24 October 2026. Source for an Exxact DGX Station listed at $94,930 in August, buyzero listing nine models from €104,800 on 24 October, 252 GB of HBM3e plus 496 GB of LPDDR5x, and two Stations linking into about 1.5 TB. https://buyzero.de/en/collections/nvidia-dgx-station-gb300\n\n[9] Minisforum US store, MS-S1 Max, read 24 October 2026. Source for $3,919 for the 128GB model. https://store.minisforum.com/products.json\n\n[10] The Register, 2 October 2026. Source for the 128GB DGX Spark at $6,950. https://www.theregister.com/systems/2026/10/02/nvidia-debuts-4999-dgx-spark-with-half-the-ram-and-storage-amid-memory-crunch/5300622\n\n[11] konstantinnikol, Kimi-K3 discussion 185 on Hugging Face, 17 September 2026. Source for 0.96 to 1.00 tokens a second on a 128GB MacBook Pro M5 Max streaming the 4-bit experts from four SSDs, with an engine from the poster's own company. https://huggingface.co/moonshotai/Kimi-K3/discussions/185\n\n[12] ComputingForGeeks, run Kimi K3 locally, updated 10 October 2026. Source for 0.10 tokens a second on a rented vast.ai server with four A100s and 2TB of RAM, GPU use between 0 and 1 per cent and its NUMA setup. https://computingforgeeks.com/run-kimi-k3-locally/\n\n[13] imec AI stack team, GPU self-hosting, updated 29 July 2026. Source for 122 tokens a second across 16 sessions on eight B300s and K3 not fitting an eight-GPU B200 node. https://aistack.imec-int.com/blog/gpu-self-hosting\n\n[14] NVIDIA, Kimi-K3-NVFP4 on Hugging Face. Source for the 1,610 GB NVFP4 build validated on eight B300 GPUs. https://huggingface.co/nvidia/Kimi-K3-NVFP4\n\n[15] OpenRouter, Kimi K3 endpoints. Source for $3 and $15 per million tokens from Moonshot and $5.50 per million out from Sail Research. https://openrouter.ai/api/v1/models/moonshotai/kimi-k3/endpoints\n\n[16] Moonshot AI, Kimi K3.1 Pro tech blog, on the Kimi blog. Source for K3.1 Pro being available on the Kimi API and its promise of the full weights by 4 November 2026. https://www.kimi.com/en/blog/kimi-k3-1-pro\n\n[17] The Next Web, Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills, 23 October 2026. Source for the release on Friday 23 October, as The Next Web reported it. https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-10-25",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.6",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "Blog",
        "url": "https://www.anchorterminal.com/blog/"
      },
      {
        "name": "Run it locally",
        "url": "https://www.anchorterminal.com/blog/run-it-locally/"
      },
      {
        "name": "Running Kimi K3 locally",
        "url": ""
      }
    ],
    "description": "Vlad Cealicu on what it takes to run Moonshot's Kimi K3, 2.8 trillion parameters with 104 billion active, yourself. The official weights are 1,561 GB, the smallest community file is 466 GB, the smallest setups shown serving it are eight-GPU B300 nodes, and what the Mac, mini PC, DGX Station and CPU server options would cost against the API.",
    "facts": [
      "Vlad Cealicu",
      "25 October 2026"
    ],
    "h1": "Can you run Kimi K3 locally? Not on a desk, and Moonshot never said you could",
    "image": "https://www.anchorterminal.com/assets/og/blog-run-kimi-k3-locally.png",
    "path": "/blog/run-kimi-k3-locally",
    "published": "2026-10-25",
    "section": "blog",
    "title": "Run Kimi K3 locally? Hardware, RAM and cost for 2.8T parameters | Anchor Terminal",
    "toc": null,
    "updated": "2026-10-25",
    "url": "https://www.anchorterminal.com/blog/run-kimi-k3-locally"
  },
  "tokens": {
    "markdown": 3150,
    "slim": 1530
  },
  "version": 1
}