Kimi K3 in the Terminal: Chinese-English Docs
Kimi K3 beside Claude Sonnet 5.5 in a docs branch
A long bilingual specification is rarely just a translation job. The Chinese section describes retry behavior, the English appendix defines error codes, and a migration note quietly changes which timeout takes precedence. The useful model is the one that keeps those relationships intact while producing something you can review in your editor.
For this comparison, put Kimi K3 beside Claude Sonnet 5.5 using the same document snapshot and task. This is a reproducible evaluation plan, not a claim that either model has already won a measured benchmark. The verdict below is about where I would start and what would make me switch.
The bilingual fixture: retries, identifiers, and conflicting limits
Build a fixture from a real repository, with secrets removed. Include Chinese design notes, an English API reference, configuration examples, and a changelog. Preserve headings and add stable line numbers before sending anything. Keep the source immutable so citations remain meaningful across both runs.
Make the task narrower than “review this document.” Ask for a terminology map, conflicting requirements, and a proposed English patch. Require every finding to cite a source section and line range. Explicitly forbid translating identifiers, changing numeric limits, or silently resolving contradictions.
- Include one Chinese requirement whose English translation reverses permission.
- Place a deprecated timeout value far from its replacement.
- Repeat one technical term with two plausible English equivalents.
- Keep JSON keys and shell flags identical across languages.
Side by side: terminology fidelity and cross-section reasoning
Kimi K3: Start with Chinese source interpretation. Ask it to distinguish normative requirements from examples and recommendations, then map each interpretation to the corresponding English passage. Judge whether it preserves conditions, exceptions, and uncertainty rather than merely producing fluent prose.
Claude Sonnet 5.5: Submit the identical request and inspect the same distinctions. Do not reward a more polished rewrite if it weakens a requirement or drops a qualification. The comparison should turn on source fidelity, not whether one answer sounds more like your team's documentation.
Kimi K3: For the long-document pass, request all references to retry limits, including references that disagree. A useful answer names the disagreement and cites both locations instead of inventing a single authoritative value.
Claude Sonnet 5.5: Apply that same requirement and compare citation accuracy. Open every cited range in your editor. A confident explanation with a nonexistent citation is a failed finding, regardless of how convincing the surrounding reasoning looks.
Side by side: patches, latency, and token accounting
Give both models a second task: return a minimal unified diff against the English document. Save each response separately, extract the patch, and run git apply --check before reviewing its semantics. Check that code fences, anchors, identifiers, and numeric literals survived unchanged unless the requested correction explicitly required otherwise.
Record elapsed time, reported usage, accepted findings, and manual repair time. Run repeated trials before drawing conclusions. Keep cold requests separate from requests reporting cache activity; repeated input does not automatically mean a cache hit. Also verify current context and output limits before attempting the complete document.
Fluxo makes this comparison convenient from a terminal: one key and one base URL, https://api.fluxorelay.com, provide access to Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok, and more. It works with Claude Code, Codex, Gemini CLI, and any OpenAI SDK. Keep the prompt and document constant, changing only the configured model.
Billing is per real token, including cache reads and writes, with no monthly fee. Compare recorded cost per accepted correction, not just cost per request. Fluxo also gives users a referral share when people they invite pay; that should stay separate from technical scoring. If this workflow fits your stack, consider creating a Fluxo account to run the fixture yourself.
The Kimi K3 verdict for Chinese-English document work
My workflow verdict: put Kimi K3 first in the evaluation queue when Chinese source interpretation dominates, but do not declare it the production winner without this fixture. Use Claude Sonnet 5.5 as the matched challenger. Choose whichever produces more verified findings and usable patches at an acceptable total cost. Fluent bilingual output is promising; traceable corrections are what earn a place in the docs pipeline.