Model Review

Kimi K3 Review: Published Weights and Dated Test Evidence

Kimi K3 has published weights under the Kimi K3 License. Our launch-period capacity failure and a later complete coding run answer different questions; neither supplies a completed frontend comparison with Fable 5. This review brings those records into one model page without promoting old rankings or a historical token rate as today's result.

Kimi K3 published weights, a dated complete run, and an unpaired Fable comparison

What the Kimi K3 model card establishes

Moonshot's official model card describes a 2.8T-parameter MoE, 104B activated parameters, and a 1,048,576-token context. These are provider specifications, not a device-fit test or an independently reproduced context result. Active parameters do not describe the complete checkpoint that a deployment must store.

The weights are published under the Kimi K3 License. The old launch note's pending-July-release language is no longer a current access statement. Read the actual terms for the intended use; this article does not give legal clearance or claim unrestricted MIT licensing.

Preserve the thinking history in a Kimi K3 tool loop

The official usage section says thinking is always enabled, with effort values low, high and max. For multi-turn calls, it requires the complete assistant message, including reasoning_content and tool_calls, to be returned in the conversation. Retaining only visible answer text can discard required state. Provider usage contract

That is an integration requirement, not a claim that every gateway or editor preserves it. Inspect the exact route, client version and sanitized tool trace. No serving engine, local weights or agent framework was installed or exercised for this revision.

Keep the incomplete launch run and later complete run separate

The launch-period report recorded eight evaluated tasks passing while repeated upstream 429s prevented parse_csv_line from completing. An unevaluated ninth task is neither a correctness failure nor a successful ninth answer. Its partial cost should not be promoted to a complete nine-task comparison.

The dated evidence archive contains a later, complete OpenRouter record. It supersedes neither the earlier request log nor the current provider contract:

Archived configurationRun / price dateScoreSuite tokens in / outDerived USD / 1,000 tasksMean latency
moonshotai/kimi-k3July 30, 2026 / July 30, 20269/91,314 / 3,538$6.3322.8 s

The stored $3/$15 per-million rates and nine-task usage produce the displayed historical amount. It is not a current Moonshot quote or an invoice for 1,000 executions. No record was rerun for this consolidation. The methodology and shared coding benchmark describe the bounded single-function tasks.

Kimi K3 vs Fable 5: the paired result is still missing

The former standalone comparison is consolidated here because it had no complete paired frontend or agent evaluation. The same archive contains anthropic/claude-fable-5 with 4/5 and excluded: true. It is not comparable to Kimi's complete nine-task row. Do not rank those fractions or silently rename Fable 5 as Fable 5.1.

Moonshot's own evaluation tables remain vendor evidence, with their stated harnesses and fallback conditions. A launch-day frontend preference ranking is not a current correctness test. Neither establishes that a generated interface submits, edits, reloads and preserves application state correctly.

QuestionEvidence retainedUnresolved decision
Does the available coding archive contain a complete Kimi run?Yes: July 30, nine short tasksRepository and frontend performance
Is there a matched complete Fable 5 row?No: the retained partial row is excludedA first-party winner claim
Does a preference vote establish working UI behavior?No functional acceptance test is supplied hereState transitions and integration correctness

A frontend decision needs a fixed acceptance test

For an actual comparison, choose one existing application change before either model starts. A useful bounded example is an editable filter that must preserve sorting and state after reload. Freeze the repository commit, dependencies, permitted files, tool access, model versions, effort settings, budget and retry rules.

Use tests for required state transitions, then inspect the same desktop and mobile viewports and keyboard behavior. Keep failed attempts, unexpected edits and review effort. A screenshot cannot override a failed submission test. Report total evaluation spend divided by accepted changes, alongside completion rate; zero accepted changes is a failure, not a fabricated finite cost.

This is a proposed evaluation, not new frontend evidence. It preserves the former comparison's practical acceptance criteria without keeping a second page that implies a completed head-to-head result.

Hosted access and self-hosting are separate decisions

For a hosted trial, obtain the current rate, exact ID and usage contract from the intended provider. For self-hosting, inspect the checkpoint, supported runtime, memory budget, licensing terms and operating costs. A lower active-parameter count does not prove a machine can load all weights, and published weights do not make serving free.

The model card links deployment references and the kimi-k3 API identifier. That is a documented option, not a first-party deployment result. Our Kimi pricing guide is separate from this dated benchmark evidence; verify any actual quote before spending.

Conclusion: choose from comparable work, not a partial comparison

Kimi K3 now has usable deployment material and a complete historical small-task record. The Fable comparison still lacks a paired result, and complex-task value remains unmeasured here. Preserve those boundaries while testing the workload you actually need.

FAQ

Are Kimi K3 weights still awaiting release?

No. The official model card provides published weights under the Kimi K3 License; deployment feasibility and use terms remain separate checks.

Did the launch run score 9/9?

The launch-period run was incomplete because one task could not be evaluated. A separate July 30 archive entry has a complete 9/9 result.

Does this review prove Kimi K3 beats Fable 5?

No. The retained Fable row is partial and excluded, so no completed paired frontend or agent result is established.

Written by

Prepared with AI assistance. This revision checks documentation, consolidates dated records and tests offline examples. It does not claim new model, integration or hardware runs.