Kimi K3 Free Online: 2.8T Open-Weight AI by Moonshot
Get Started Free

Kimi K3

Current version: K3 updated

Largest open-weight model: 2.8T params, 1M context

Category
Language
Modality
Text · Reasoning · Tool use
Context
1M
Released
Jul 2026
Run this model Estimate cost
Update

What's new in Kimi K3

  • 2.8T parameters versus 1T in K2.6 — the largest open-weight model released
  • Context window of 1,000,000 tokens, enough for a full repository
  • First place in Frontend Code Arena at 1679, ahead of Claude Fable 5
  • GDPval-AA v2 score of 1687, third behind Fable 5 Max and GPT-5.6 Sol Max
  • Weights published for unrestricted download
Strengths

What it's the best tool for

  • 2.8T parameters — the largest open-weight model, downloadable without restrictions
  • 1,000,000-token context loads a large repository in full
  • First place in Frontend Code Arena, ahead of Claude Fable 5 in blind developer testing
  • Strong agentic behaviour: tool use, debugging, iteration on logs and tests
  • Cached input costs roughly ten times less than uncached
  • Easier to self-host than comparable frontier models
Limitations

When to reach for something else

  • Roughly five times more expensive per token than K2.6
  • Moonshot temporarily throttled access after launch due to demand
  • The full technical report trailed the weight release
  • As a Chinese model it is reserved on politically sensitive topics
  • Self-hosting demands serious hardware despite the efficiency gains
Sample output

How Kimi K3 responds

Prompt
Turn this repository into a working prototype: fix the failing tests, add the missing endpoints and show me what is left to do.
Kimi K3
Loaded the full repository, 62 files. Started with the tests. 7 of 41 were failing. Five shared one cause: after the ORM migration the fixtures kept the old signatures. Fixed those as a batch. The other two were real pagination bugs — offset counts from zero while the cursor counts from one. Patched in repository/paginate.py. Added three endpoints present in the spec but missing from the code: DELETE /items/{id}, PATCH /items/{id}/status, GET /items/export. Still open: file uploads have no validation and the billing module has no test coverage — that is the next pass.
Where teams use it

Four scenarios where it pays for itself

01
Agents over large repositories
A 1M context holds the whole project so the model keeps cross-file context
02
Frontend work
First in Frontend Code Arena, rated above Claude Fable 5 by developers
03
Log-driven debugging
Reads traces, tests and screenshots, iterating until it works
04
Self-hosting
Open weights let you run it on your own hardware without vendor lock-in
About model

More about Kimi K3

Kimi K3 — the largest open-weight model available

Kimi K3 from Moonshot AI runs 2.8 trillion parameters, making it the biggest open-weight model released publicly. The weights are available for unrestricted download — anyone with the hardware can self-host it.

What K3 is good at

The context window holds 1,000,000 tokens, enough to load a large repository in full so the model keeps track across files. It was built for agentic work: navigating big codebases, calling tools, debugging, and iterating against logs, tests, images and runtime feedback.

On benchmarks it trades blows with proprietary flagships. GDPval-AA v2: 1687 points, third overall behind Claude Fable 5 Max (1815) and GPT-5.6 Sol Max (1747.8), ahead of Claude Opus 4.8 (1600). AA-Briefcase: 1527, second place, beating GPT-5.6 Sol Max (1495). And in Frontend Code Arena it took first place at 1679 — ahead of Claude Fable 5 in blind developer testing.

Pricing

Open weights do not mean a free API. Through providers, K3 runs about $2.90 per million input tokens and $14 per million output — a real step up from K2.6. Cached input drops to roughly $0.30 per million, so long conversations with repeated context cost far less than the headline rate suggests.

Access on NetRoom

Demand after launch was heavy enough that Moonshot temporarily restricted access to the model. On NetRoom it is wired up directly — runs in the browser, no separate account, billed per token actually used.

Who it's for

Teams running agents across large repositories, anyone who values open weights, and anyone who wants a fallback if a proprietary vendor changes the rules.

Versions

Version history of Kimi K3

Version Date What changed
K3 current
  • 2.8T parameters versus 1T in K2.6 — the largest open-weight model released
  • Context window of 1,000,000 tokens, enough for a full repository
  • First place in Frontend Code Arena at 1679, ahead of Claude Fable 5
  • GDPval-AA v2 score of 1687, third behind Fable 5 Max and GPT-5.6 Sol Max
  • Weights published for unrestricted download
K2.6
  • 1T MoE parameters with 32B active per token
  • Agent Swarm scaling to 300 domain-specialised sub-agents
  • Up to 4,000 coordinated steps in a single autonomous run
  • Released under a Modified MIT licence

Try Kimi K3
right now

Free access to basic models. No card, no obligations.