Kimi K3: 180K downloads, 340 fine-tunes, 72 hours

Moonshot AI released the largest open-weight model ever on July 27, 2026, and for three days the numbers moved like a product launch, not a research release: 180,000 Hugging Face downloads, 340+ community fine-tunes, and a computing crunch so bad the company paused new consumer subscriptions.

The story has two halves. The Western developer community treated K3 as an event. The Chinese side spent the same week arguing about a 3.5x price increase and a capacity crisis. Both halves are true, and both are worth compiling properly.

What shipped

Kimi K3 is a 2.8-trillion-parameter MoE, 104B active per token, with a 1M token context window and native vision [1]. The architecture is the unusual part: KDA, a 3:1 mix of linear and full attention layers, plus attention residuals, cuts KV cache memory by about 75% and speeds up long-context decode by up to 6.3x [1][2]. It activates 16 of 896 experts per token and claims 2.5x the scaling efficiency of K2 [1].

The weights went up on Hugging Face on July 27 under a Modified MIT license, alongside the technical report and Moonshot's training infrastructure (MoonEP, FlashKDA, AgentEnv) [3][4]. Modified MIT means commercial use, fine-tuning, and private deployment are allowed; training data and code stay private [5].

Self-hosting is not for hobbyists. The full weights are 1.56TB across 96 safetensors, and practical serving wants 8x H100 or equivalent with tensor parallelism [6]. The native MXFP4 format means no post-hoc quantization is needed, but you still need a cluster.

The 72-hour numbers

vLLM shipped Day-0 support with a production config on release day [8]. That's the detail that separates K3 from earlier open releases: the serving stack was ready before the hype peaked.

Where it ranks

The numbers are mixed, and we're labeling each by nature.

• Frontend Code Arena: 1679, first place, ahead of Claude Fable 5 (1631) and GPT-5.6 Sol (1618). It's marked Preliminary, with fewer votes (1757 vs Fable 5's 2505) [10]. Third-party eval.
• SWE Marathon: 42.0, the highest published score [11]. Third-party eval, per vendor release materials.
• Terminal-Bench 2.1: 88.3, second only to GPT-5.6 Sol (88.8) [2]. Third-party eval, per vendor release materials.
• Artificial Analysis Intelligence Index: 57.1, third of 189 models, about 2-3 points behind Fable 5 (59.9) and GPT-5.6 Sol (58.9) [12]. Third-party eval.
• DeepSWE: 67.5, below Fable 5 (70.0) and Sol (73.0) [11]. Third-party eval.

Moonshot's own release notes say K3 still trails Fable 5 and GPT-5.6 Sol overall but beats everything else it tested [1]. That's a vendor statement, so we mark it vendor-reported. The honest summary: K3 wins some coding races, loses others, and sits 2-3 points below the frontier on the aggregate index.

Pricing: the expensive open model

K3 is priced at $3 per million input tokens and $15 per million output tokens, with cache hits at $0.30 [9][12]. In RMB that's 20/100 yuan, about 2.1x and 2.7x Kimi K2.6, and roughly 10-12.5x MiniMax M3 [13]. Moonshot's enterprise lead said it plainly: open source doesn't have to be cheap, SOTA capability deserves SOTA pricing [13].

Then demand hit. Within 48 hours of launch, request volume approached cluster capacity and Moonshot paused new consumer subscriptions on July 19 to protect existing users [13][14]. Bloomberg reported daily sales up at least 6x after launch, with June ARR around $300 million and an IPO on Hong Kong's exchange in motion [14][15].

Our take

The 180K downloads and 340 fine-tunes are the real story, and they happened because the tooling was ready: vLLM Day-0, Ollama one command, cloud mirrors. K3 got the treatment that DeepSeek got a year earlier, except faster, because the ecosystem now knows what to do with a frontier open weight.

The demand crush is the other half. A model that hits capacity in 48 hours and pauses subscriptions is a supply problem, not a demand problem, and it tells you the Chinese open-weight boom is hitting a wall that has nothing to do with model quality: compute.

The political noise around K3 (a White House "industrial distillation" accusation with no evidence released, then a reported exemption for Chinese open weights from US safety tests a week later [16]) is best treated as weather. What survives weather is infrastructure: the 1M context, the MXFP4 weights, the Day-0 serving stack. We compiled what we can verify and left the rest in the archive.

Risks and what we're watching

• Capacity. Moonshot reopened subscriptions in batches, and the company says it's expanding compute [14]. Watch whether pricing holds or the cache-hit price moves.
• Self-hosting cost. 1.56TB of weights and 8x H100 is out of reach for most teams; community quantized versions exist but we haven't verified their quality.
• The "most expensive Chinese model" label [13][17] is a real adoption filter for API users, even with cache hits at $0.30.
• Fine-tune quality: 340+ derived models in 72h is volume; we're tracking which ones show up in actual production.

Bottom line

Kimi K3 is the largest open-weight model ever released, with the strongest day-one tooling story we've tracked, a genuine coding-adjacent lead on Arena, and a price tag that broke the "open source is cheap" assumption. The downloads and the fine-tunes say the market accepted the trade. The paused subscriptions say compute, not demand, is the constraint. Both facts are on the leaderboard.

Same number on our leaderboard → Kimi K3 row

Sources

1. IT Home (K3 release, architecture, vendor positioning), ithome.com/0/982/259.htm (published 2026-07-27, captured 2026-08-06, media report)
2. NodeMini blog (KDA details, Terminal-Bench 2.1 88.3), nodemini.com/zh/blog/2026-kimi-k3-open-source-28t-parameter-review.html (published 2026-07-16, captured 2026-08-06, community)
3. Guangzhou Daily (weights + tech report + infra release), huacheng.gz-cmc.com/pages/2026/07/28/6370587b2a104859af17da3aa866fc36.html (published 2026-07-28, captured 2026-08-06, media report)
4. OSChina (open-source day: weights, tech report, Moonshot infra), my.oschina.net/u/9487999/blog/19722953 (published 2026-07-17, captured 2026-08-06, media report)
5. CNBlogs (Modified MIT license terms), cnblogs.com/tutudatu/p/21960857 (published 2026-07-27, captured 2026-08-06, media report)
6. OpenModelMap (1.56TB / 96 safetensors / 8x H100), openmodelmap.com/kimi-k3/zh (published 2026-08-01, captured 2026-08-06, community)
7. CSDN (72h stats: 180K downloads, 340+ fine-tunes, frameworks 48h), blog.csdn.net/xyghehehehe/article/details/163353310 (published 2026-08-05, captured 2026-08-06, media report)
8. vLLM Blog (Day-0 support, production config), vllm.ai/blog/2026-07-27-k3 (published 2026-07-27, captured 2026-08-06, official)
9. Ollama (kimi-k3:cloud, pricing, downloads), ollama.com/library/kimi-k3 (published 2026-08-01, captured 2026-08-06, official)
10. Jiemian (Frontend Code Arena 1679, Preliminary flag), jiemian.com/article/14786633.html (published 2026-07-17, captured 2026-08-06, third-party eval, via media)
11. QbitAI (SWE Marathon 42.0, DeepSWE 67.5), qbitai.com/2026/07/455179.html (published 2026-07-20, captured 2026-08-06, third-party eval, via media)
12. MACGPU Blog (AA Intelligence Index 57.1 #3/189), macgpu.com/zh/blog/2026-0722-kimi-k3-kaiyuan-quanzhong-fabu.html (published 2026-07-22, captured 2026-08-06, third-party eval, via community blog)
13. Tencent News / Jiupai (pricing vs K2.6, M3 anchor, capacity pause), news.qq.com/rain/a/20260724A0A6XX00 (published 2026-07-24, captured 2026-08-06, media report)
14. OpenAI Hub (timeline: 48h capacity, July 19 pause, batch reopening), openai-hub.com/news/1148 (published 2026-07-19, captured 2026-08-06, media report)
15. QbitAI / Bloomberg (daily sales 6x, ARR $300M, HK IPO), qbitai.com/2026/07/455179.html (published 2026-07-20, captured 2026-08-06, media report of Bloomberg reporting)
16. OSChina / Bloomberg (White House spares Chinese open weights from US safety tests), oschina.net/news/488173/china-s-open-weight-models-to-be-spared-us-tests-us-firms-told (published 2026-08-05, captured 2026-08-06, media report of Bloomberg reporting)
17. Juejin (domestic criticism: most expensive, slow), juejin.cn/post/7663685196478611506 (published 2026-07-18, captured 2026-08-06, community)

Disclosure: no sponsors, no affiliate links. Compiled from public sources with capture dates; we did not run K3 ourselves. Vendor-reported and preliminary scores are labeled as such.