Moonshot AI · in Borg

Kimi K3

Kimi K3 is Moonshot AI’s flagship K3 generation, a larger open-weight model aimed at demanding engineering and knowledge-work agents rather than the narrower, lower-priced K2.7 Code.

Reasoning1,048,576 tokenstext, image, video

1,048,576 tokensContext window

8627 msMedian time to first token (gateway proxy)

18.1Output tokens per second (whole turn)

$3 / $15Vendor API price in / out · per 1M

Included in Borg

Download Borg Borg for Business

At a glance

Specification	Value
Provider	Moonshot AI
Family	Kimi
Release date	2026-07-16 Vendor source
Context window	1,048,576 tokens Vendor source
Input modalities	text, image, video
Reasoning	Yes
Open weights	Yes
Borg lane ids	`k3`, `k3-256k`, `moonshotai/kimi-k3`
Median time to first token (gateway proxy) on Borg	8627 ms
Median tokens per second (end to end) on Borg	18.1 output tokens per second

Not published by the vendor: output modalities, thinking modes, knowledge cutoff, fusion roles.

What Kimi K3 is good for

Kimi K3 is Moonshot AI’s flagship K3 generation, a larger open-weight model aimed at demanding engineering and knowledge-work agents rather than the narrower, lower-priced K2.7 Code. Its long context can accommodate extensive repositories, documentation and tool histories; image and video understanding lets it inspect visual evidence while working on interfaces, designs or research. It always reasons, with adjustable effort for trading latency and token usage against difficult problem solving. Compared with K2.7 Code, it is the broader choice for mixed research, visual analysis and long-running projects, but its direct API token rates are higher. It is particularly useful when an agent must coordinate terminal operations and reconcile information spread across many files. Borg includes K3 in its subscription; Borg Fusion can route work to it, or users can choose its lane directly.

long-horizon coding across large repositories
research agents synthesizing extensive source material
visual debugging of interfaces and design work
tool-driven engineering with persistent context

How Borg uses it

Kimi K3 is available as a subscription catalog lane in Borg. The catalog does not publish which Fusion profiles route to this model, so no routing role is claimed here. You can pick it directly by setting the OpenAI-compatible gateway model field to a listed lane id. Check the current catalog for availability before building a workflow around a lane.

model: "k3"

Measured on Borg

Metric	Median	95th percentile	Samples
Time to first token (gateway proxy)	8627 ms	42705 ms	7633
Tokens per second (end to end)	18.1 output tokens per second	41.4 output tokens per second	152924

Time to first token (gateway proxy)

Lower is fasterFleet median 3971.5 ms

Kimi K38627 ms

Output tokens per second (whole turn)

Higher is fasterFleet median 31.9 tokens/s

Kimi K318.1 tokens/s

Time to first token (gateway proxy): time until the first streamed response event reaches the gateway; our closest available measure of time to first token
Tokens per second (end to end): tokens per second, end to end: provider-reported output tokens divided by the whole turn's elapsed time, including queueing and reasoning; not a streaming per-token rate

Completion speed covers the whole turn including queueing and reasoning time; it is not a streaming tokens-per-second figure.

Measured on Borg: SearchAtlas AI Gateway production request logs. 30-day window ending 2026-10-08; 183015 requests. Minimum 100 samples per published request metric.

Benchmarks

Benchmark	Vendor-reported score	Source
GPQA Diamond	93.5	Vendor source
DeepSWE	67.5	Vendor source
ProgramBench	77.8	Vendor source
FrontierSWE	81.2	Vendor source
SWE-Marathon	42	Vendor source

Pricing

Included in Borg subscription. Vendor API list price: $3 input / $15 output per 1M tokens. Vendor price source.

Compare subscription and API pricing · Cost Savings Calculator

FAQ

What is Kimi K3?

Kimi K3 is listed in Borg's subscription catalog as k3. Kimi K3 is Moonshot AI’s flagship K3 generation, a larger open-weight model aimed at demanding engineering and knowledge-work agents rather than the narrower, lower-priced K2.7 Code.

Is Kimi K3 in Borg?

Yes. Kimi K3 is listed in Borg's subscription catalog under the lane k3. You can choose the lane directly, or explore the available Fusion profiles. Availability and model details can change; check the current catalog before deployment.

What context window does Kimi K3 have?

The published context window is 1,048,576 tokens. Consult the cited sources for details and current limits. Vendor source

Is Kimi K3 good for coding?

Vendor-published evaluations appear in the benchmarks section; their settings may differ from your tasks. Its listed uses include long-horizon coding across large repositories, research agents synthesizing extensive source material, visual debugging of interfaces and design work, tool-driven engineering with persistent context. Try the lane for your own tasks, and compare the output with your requirements before choosing a default.

How fast is Kimi K3 on Borg?