When are Kimi K3 open weights released?

Moonshot AI announced on July 16, 2026 at WAIC 2026 in Shanghai that Kimi K3's full 2.8-trillion-parameter weights will be released on July 27, 2026 under a Modified MIT license. Until then, K3 is only accessible via the Kimi app, Kimi Code, and the Kimi API. On July 27, the weights become downloadable from Hugging Face and Moonshot's own registry, enabling self-hosting and fine-tuning.

What does Kimi K3 cost via API?

Kimi K3 API pricing (July 2026): $3.00 per million input tokens (cache-miss / fresh), $0.30 per million input tokens (cache-hit), and $15.00 per million output tokens. All rates apply across the full 1-million-token context window. This is significantly cheaper than Claude Opus 4.8 ($5/$25), Fable 5 ($10/$50), and GPT-5.6 Sol, but comparable to Sonnet 5 output pricing. Kimi Code (the Claude Code-style CLI) uses the same API rates.

What hardware do I need to self-host Kimi K3?

Kimi K3 is a 2.8T-parameter Mixture-of-Experts model with 16 of 896 experts active per token (~50B active parameters). Rough minimum hardware for FP8 inference: 8× H200 141GB (about $250K-$300K hardware cost), or 8× MI325X. For MXFP4 quantized inference (Moonshot ships MXFP4 checkpoints), you can fit K3 on 4× H200 or 4× MI325X, dropping the cost to roughly $130K. For batch throughput and low latency, plan on 8-16 accelerators plus fast NVLink/InfinityFabric interconnect. Small teams should self-host via a GPU cloud (Lambda, RunPod, Together, Fireworks) rather than owning hardware.

When does self-hosting Kimi K3 beat using the API?

The API break-even for self-hosting is roughly 100-300 million tokens per day of steady traffic on a rented 8× H200 instance ($30-$60/hour on Lambda / RunPod). Below that, the API at $3/$15 per MTok wins on total cost. Above that, self-hosting wins economically — and self-hosting also wins any time you need data residency (Chinese model on Chinese cloud, or fully on-prem), custom fine-tuning, or predictable latency. Cursor, Continue, and Cline all support Kimi K3 via the API today; self-hosted routing will land within days of the July 27 weights drop.

Quick Answer

Kimi K3 Self-Hosting Guide: Open Weights July 27 (2026)

Published: July 18, 2026

Kimi K3 Self-Hosting Guide: Open Weights July 27 (2026)

On July 16, 2026, at WAIC 2026 in Shanghai, Moonshot AI released Kimi K3 — a 2.8-trillion-parameter Mixture-of-Experts model with a 1-million-token context window — and promised full open weights by July 27, 2026 under a Modified MIT license. That makes K3 the largest open-weight model ever released, and one of the first frontier-tier models where self-hosting is economically viable for teams above ~200M tokens/day.

Here is what you need to know: pricing today, weights availability, hardware requirements, and when self-hosting beats the API.

Last verified: July 18, 2026

What Was Announced

Date: July 16, 2026, at WAIC 2026 (Shanghai).
Model: Kimi K3 — 2.8T total parameters, MoE with 16 of 896 experts active per token (~50B active).
Context: 1 million tokens, native multimodal (vision).
Weights release: July 27, 2026 (Modified MIT license). Hugging Face + Moonshot registry.
API pricing (available now via Kimi API):
- Input (cache-miss): $3.00 / MTok
- Input (cache-hit): $0.30 / MTok
- Output: $15.00 / MTok

Moonshot’s own benchmarks claim K3 beats Fable 5 and GPT-5.6 Sol on frontend code arena benchmarks, and beats Opus 4.8 and GPT-5.5 on its internal coding/agentic evaluation suite — while acknowledging K3 generally lags Fable 5 and Sol on overall difficulty.

The Pricing Comparison (July 18, 2026)

Model	Input $/MTok	Output $/MTok	Context	Weights
Kimi K3	3.00	15.00	1M	Yes (Jul 27)
Claude Sonnet 5	5.00	25.00	1M	No
Claude Opus 4.8	5.00	25.00	1M	No
Claude Fable 5	10.00	50.00	1M	No
GPT-5.6 Sol	not yet public	not yet public	400K	No
Gemini 3.5 Pro	15.00	60.00	2M	No
MiniMax M3	0.60	2.40	1M	Yes (MIT)
DeepSeek V4 Pro	0.44	0.87	1M	Yes (MIT)

Read: K3 is priced to undercut Fable 5 heavily, undercut Gemini 3.5 Pro by 5×, and match Sonnet 5. It is priced above MiniMax M3 and DeepSeek V4 Pro — the trade-off is claimed higher performance at the top of the Chinese open-weight range.

Hardware for Self-Hosting K3

K3 is a big model. Rough sizing (FP8 vs MXFP4 checkpoints):

Setup	GPU	GPUs	Purpose	Approx cost (hardware)
FP8 minimum	H200 141GB	8	Single-user, medium throughput	~$260K
FP8 minimum	MI325X 256GB	8	Single-user, medium throughput	~$220K
MXFP4 quantized	H200 141GB	4	Single-user, decent throughput	~$130K
MXFP4 quantized	RTX 6000 Blackwell 96GB	8	Workstation prototype	~$45K
Serving cluster	H200 141GB	16	Multi-tenant / batch	~$520K

For teams: rent don’t buy. Lambda, RunPod, Together, Fireworks, and DeepInfra will all host K3 within a week of the July 27 weight drop. Expect $30-$60/hour for an 8× H200 instance. For $3-5K/month of steady rental you can serve 300-500M tokens/day.

For hobbyists: wait a week after July 27 for community MXFP4 quantizations that fit on 4× RTX 6000 Blackwell or 2× MI325X. Expect community llama.cpp / vLLM / SGLang integrations same week.

When Self-Hosting Beats the API

API wins when:

Traffic < ~100M tokens/day (steady).
You need burst capacity without pre-provisioning.
You care about vendor-managed uptime.
You are prototyping / evaluating.

Self-hosted wins when:

Traffic > ~300M tokens/day steady.
Data residency matters (e.g. cannot send to Chinese API; or must stay in EU / on-prem).
Custom fine-tuning is a competitive advantage.
Predictable latency requirements.
You want to swap between K3, MiniMax M3, DeepSeek V4 Pro, and GLM-5.2 based on task cost — cheaper to keep infra warm than run four API bills.

Break-even math (rough):

8× H200 rental: ~$40/hour × 720h = $28.8K/month.
Equivalent API cost at $3/$15: with 30/70 input/output mix, average $11.40 / MTok. $28.8K buys 2.53 billion tokens/month = 84M tokens/day.
Break-even: ~85-100M tokens/day steady. Above that, self-hosting wins; below that, the API wins.

Add fine-tuning, data residency, or predictable latency as multipliers that make self-hosting attractive at lower volumes.

Tool Integration Status (July 18, 2026)

Already working with K3 via API:

Cursor (add as custom model, OpenAI-compatible endpoint).
Continue.dev (add via config.yaml).
Cline / Roo Code (custom OpenAI-compatible provider).
Aider (--model kimi/k3).
Kimi Code — Moonshot’s own Claude-Code-style CLI, native.

Expected within days of July 27:

vLLM native support.
SGLang native support.
Ollama community quantizations.
LM Studio community quantizations (MXFP4).
Together, Fireworks, DeepInfra hosted endpoints.

The Modified MIT License Question

Kimi K3 ships under a Modified MIT license — meaning some conditions attach vs pure MIT. Watch for these clauses in the actual license text (release July 27):

Prohibited uses (military, some critical infrastructure, or specific downstream products).
Attribution requirements (must credit Moonshot for downstream products above a size threshold).
Redistribution restrictions (some frontier open-weight licenses require notifying the vendor for large redistributions).

Most enterprise use will be fine, but review the license before shipping a commercial product built on K3.

Sub-Questions Buyers Are Asking

Is K3 actually good enough to displace Claude Sonnet 5 for coding? Not yet on independent benchmarks. Moonshot’s numbers are internal. Wait for artificialanalysis.ai and Kilo/BenchLM independent scores over the next 7-10 days. For coding today, Claude Sonnet 5 and GPT-5.6 Sol remain the highest-confidence picks.

Can I use K3 in China-restricted contexts (US federal, EU financial)? K3 is a Chinese model — most US federal contracts and some EU regulated financial contexts will not approve it, regardless of open weights. Self-hosting on your own EU/US infra addresses data residency but not model-provenance policies.

Is Kimi K3 open enough to fine-tune? Yes, once weights land. Community LoRA and full fine-tuning pipelines will appear within a week. For MoE fine-tuning at this scale, use torchtune, unsloth (once K3 support lands), or MegaBlocks. Expect $50-200K in compute for a serious full fine-tune.

How does K3 compare to DeepSeek V4 Pro on cost? DeepSeek V4 Pro is roughly 7× cheaper on API ($0.44/$0.87 vs $3.00/$15.00). If your task doesn’t need K3’s frontier capability, V4 Pro is a stronger open-weight budget pick. Use K3 when you need Fable-5-adjacent performance at 30% of Fable 5’s price.

Bottom Line

Kimi K3 is the first open-weight model where “self-host for real” becomes an obvious CTO conversation, not a hobbyist debate. With frontier-adjacent quality, 1M context, native vision, and a real MIT-family license coming July 27, K3 changes the math for teams doing more than 100M tokens/day of steady traffic.

For most teams: use the API today ($3/$15). Wait for independent benchmarks before making K3 your default coding model. Plan self-hosted capacity for July 30 onward if data residency, fine-tuning, or high steady volume applies. And keep DeepSeek V4 Pro and MiniMax M3 in the mix — the Chinese open-weight ecosystem is now genuinely competitive top to bottom.

Sources

Moonshot Kimi K3 blog: kimi.com/blog/kimi-k3
OpenRouter Kimi K3 listing: openrouter.ai/moonshotai/kimi-k3
Hugging Face K3 model card: huggingface.co/blog/ResterChed/kimi-k3-model-overview-mxfp4-quantization-open-wei
Tom’s Hardware coverage: tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3
artificialanalysis.ai independent benchmarks: artificialanalysis.ai/models/kimi-k3