Kimi K3 — agentflows stackKimmy works here.

Designation

The Worker

Default by designation.

Kimi K3 is the default worker of the agentflows stack — the agent that gets the ticket first. It reviews, implements, plans, and orchestrates. Everything it ships passes through structural human gates.

Economics

Cost

Marginal cost: approximately zero.

K3 runs on a flat-rate subscription — quota-limited, never metered per call. Send it another task; the invoice doesn't move.

Memory

Context

262K tokens of working memory.

One window, wide enough for the repo, the diff, and the argument about the diff.

Recall

Review

100% recall. Obvious bugs and subtle ones.

In repo evals, K3 caught every bug — the loud ones and the quiet ones — tying Claude Sonnet at ~1/7 the cost.

Throughput

Execution

110/110 tests. 8 turns.

The most turn-efficient implementer measured: 8 turns to a passing build, versus 20 for legacy and 15 for Sonnet.

Tradeoff

Planning

Tied with Sonnet at 4/5 — with an asterisk.

Judges scored K3's plans equal to Sonnet's: stronger on concision, weaker on verification. It writes the plan well. It checks the plan less well. That's what the gates are for.

Command

Orchestration

Four agents. 3 P0. 8 P1. Every citation real.

In an AgentSwarm run, 4 read-only K3 subagents surfaced 3 P0 and 8 P1 issues — 6/6 citations verified. K3 also conducts MiniMax-M3 cross-checkers and Claude councils, all under structural human gates.

Measured

Field data

262K

context window

~$0

marginal cost · flat-rate, quota-limited

100%

bug recall, obvious & subtle

~1/7

cost vs. Claude Sonnet in review

110/110

tests passed, coding eval

8

turns to passing build · vs. 20 legacy, 15 Sonnet

4/5

planning, judge-tied with Sonnet

−46% / +11%

caveman words, condensing / generating

3P0 + 8P1

found by 4 read-only K3 subagents

6/6

citations verified