Kimi K3 & K3.1: Release Date, Weights and Local Specs

Quick answer

If you came here to find out which GPU runs Kimi K3: none on its own. Even the most aggressive GGUF, Unsloth's 466 GB UD-Q1_0, is larger than any single card or most workstations. K3 is built for multi-GPU servers; Moonshot's model card recommends vLLM, SGLang or TokenSpeed to serve it, and the 1.56 TB official weights are already 4-bit, so there is no 8-bit-to-4-bit saving left to take.

The practical playbook: use the K3 or K3.1 API if you want the model, and self-host from the tier below it. The Kimi K2 line is the runnable branch of the family, and GLM-5.2 is a frontier-class open model that is far more practical to self-host.

Kimi K3.1 Pro: release date and weights

Kimi K3.1 Pro rolled out on October 23, 2026, according to Moonshot's launch post, on Kimi.com, the mobile apps, Kimi Work, Kimi Code and the Kimi API, with a 1-million-token context. Moonshot calls it its most capable “3T-class open-weight” model and says the full weights are due by November 4, 2026.

One detail matters for anyone planning to self-host: Moonshot says it identified and removed cyber-related data from both training stages of the open-weight release, and reports its internal offensive-security test score dropping as a result. The open K3.1 weights will therefore not be identical in behaviour to the API model.

Moonshot has not published a parameter count for K3.1 Pro beyond “3T-class”, so plan on hardware at least as large as K3's until the files are on Hugging Face and their size is known.

Kimi K3 specs at a glance

Source: the moonshotai/Kimi-K3 model card. One naming note: you'll occasionally see this model called "Kimi V3." There is no Kimi V3 — the real name is K3, the successor to the K2 line.

Release date & license

K3 launched July 16, 2026 through Moonshot's API and app — see Moonshot's official Kimi K3 announcement. Moonshot had promised open weights by July 27, and kept the date: the full weights, the technical report and a KDA prefix-caching implementation for vLLM were published that day.

The licence turned out to be its own text, the Kimi K3 License, rather than the modified MIT of the K2 line. It grants the usual MIT rights, then adds two conditions: a company running a model-as-a-service business with more than US$20 million revenue over 12 months must sign a separate agreement before commercial use, and products with more than 100 million monthly users or US$20 million monthly revenue must display “Kimi K3”. Internal use is exempt from both.

The vLLM integration means multi-node operators had a serving path on day one. KDA is a new attention design, so llama.cpp support arrived through Unsloth's fork rather than a release of mainline llama.cpp, and Ollama offers K3 only as a cloud model.

Can you run it locally? Only on very big iron

The arithmetic, with real file sizes now that the weights are out:

Kimi K3 GGUF sizes and the hardware they need

Sizes from unsloth/Kimi-K3-GGUF (total of the split files). You need at least the file size in combined RAM and VRAM, plus room for the context cache.

Unsloth's README says to run these with its llama.cpp PR fork or Unsloth Studio. NVIDIA publishes nvidia/Kimi-K3-NVFP4 for Blackwell-class servers.

API pricing

For most people the API is the practical way to use K3. Launch pricing (July 2026), flat across all context lengths:

At launch Moonshot reported over 90% cache hits on coding workloads, which would put most input at the $0.30 rate. Output is where K3 was premium-priced; for comparison, DeepSeek V4 lists far cheaper output. Check Moonshot's platform for current prices.

Benchmarks — Moonshot's claims, labeled as such

The model card's evaluation tables are Moonshot's own runs, mostly with its Kimi Code harness, mixed with scores it cites from Artificial Analysis and public leaderboards as of July 23, 2026. Early independent API testing includes Simon Willison's write-up. Treat vendor tables accordingly.

• →Moonshot itself says K3 trails Claude Fable 5 and GPT-5.6 Sol overall. Unusual candor for a launch post, and a point in their credibility column.
• →Moonshot's numbers have K3 beating Claude Opus 4.8 and GPT-5.5 on several coding and agentic benchmarks — for how those claims compare to independently-scored models, see the SWE-bench Verified leaderboard explainer.
• →K3.1 Pro: Moonshot's launch post says its offensive-cyber capability remains behind frontier proprietary models; independent scores are still arriving.

Frontend Code Arena at K3's launch (July 2026), as reported in Moonshot's launch coverage:

Arena scores move every week. For the current crowd-voted ranking, check where K3 ranks on the Arena (LMArena) leaderboard.

What you can run instead

If K3 sent you here dreaming of frontier-class local AI, here's the realistic ladder, from extreme to actually-attainable:

Kimi K3 vs K2: what actually changed

K3 is a scale jump, not a tune-up: 2.8T total and 104B active parameters against about 1.03T for Kimi K2.6 (April 2026) and K2.7-Code (June 2026), a 1M-token context, native vision, and the new KDA attention. Moonshot says the Stable LatentMoE design gives roughly 2.5x better scaling efficiency than K2. The licence changed too: the K2 line uses modified MIT, K3 its own Kimi K3 License.

For local users the comparison is simpler than any spec sheet: the K2 line is the one you can run. K2.7-Code's 2-bit GGUFs are 318-340 GB; K3's smallest 2-bit file is 711 GB and its full weights 1.56 TB. If you're choosing a Kimi model to self-host, start with K2.7-Code for coding or K2.6 for general use.

Bottom line

Kimi K3 is a real open-weight frontier release — 2.8T parameters, a 1M-token context window, weights you can download today — but at 1.56 TB, with GGUFs starting at 466 GB, it is a server model. K3.1 Pro is the newer model to watch, with weights due by November 4, 2026.

The practical playbook: use the API if you want K3 or K3.1, and self-host from the tier below. That means Kimi K2.7-Code on a 384 GB+ machine, or GLM-5.2 for frontier-class capability you can actually run.

FAQ