Hy4 vs Kimi K3.1 Pro: 3x faster, and it explains itself

Tencent Hy4 is the faster of the two by a wide margin, and the more talkative. On the same 41 prompts, its median response time was 48 seconds to Kimi K3.1 Pro's 153, and it finished first on 38 of them. It also wrote about 1.7x as much text, and nearly all of the difference is commentary around the answer rather than a longer answer.

Both are reasoning models from Chinese labs, released 12 days apart with roughly 1M-token context windows, and Tencent's launch post claims Hy4 beat Kimi K3.1 Pro in a finance evaluation. The spec sheets don't separate them, so we compared what they actually return.

The numbers

We used the 41 text prompts on FastMetal's compare pages: 20 in Japanese and 21 in English, covering debugging, code review, SQL optimization, business email, translation, meeting-notes summaries, structured data extraction and word problems. The two image-generation prompts were left out, since both are text models. Same prompts, default settings, no reasoning-effort changes. Kimi K3.1 Pro ran on October 27, 2026 and Hy4 on November 6, 2026, both through FastMetal's API. Response times are what we measured on our route and will vary with load; 41 prompts is a small sample, so read these as tendencies.

The gap widens on prompts that need long thought: English code review took Hy4 110 seconds and Kimi K3.1 Pro 560. On short ones it nearly closes; counting the r's in "strawberry" took 4 and 11 seconds.

Where Hy4's extra text comes from

Put the answers side by side and the pattern is obvious. Kimi K3.1 Pro returns what you asked for. Hy4 returns that, then explains it.

• Rewriting a dense Japanese paragraph about the CAP theorem, Kimi K3.1 Pro returned the rewritten paragraph (180 characters). Hy4 returned the rewrite plus a bulleted list of what it changed and why, including a correction to how the original framed CAP (622 characters).
• Translating a Japanese product review into English, Kimi K3.1 Pro returned the translation. Hy4 returned two translations, a casual one and a more formal one, with notes in Japanese on how it handled the nuances. Its casual version ("If I had to nitpick… but hey, that's a pretty good problem to have") reads more like a real Amazon review than Kimi K3.1 Pro's.
• Writing a Japanese business email, Hy4 added fill-in placeholders and a line on how to use the template.

For a person choosing between options, that commentary is useful. For a pipeline that feeds the output to the next step, it is text you have to strip, and you pay for it as output tokens.

The extreme case: structured extraction

Asked to pull company and financial details from a news article into JSON, Kimi K3.1 Pro returned 2,508 characters. Hy4 returned 25,243. It nested every fact with its raw wording, a normalized value and a note ("today" became an object recording that no absolute date was given). Thorough, and not the shape anyone downstream expects.

The fix is to hand Hy4 a schema. Before launch we confirmed that Hy4 enforces strict json_schema on FastMetal, even when the schema has nothing to do with the prompt:

curl https://api.fastmetal.ai/v1/chat/completions \
  -H "Authorization: Bearer $FASTMETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hy4",
    "messages": [{"role": "user", "content": "Extract the acquirer, the target and the deal value from this article: (article text)"}],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "deal",
        "strict": true,
        "schema": {
          "type": "object",
          "properties": {
            "acquirer": {"type": "string"},
            "target": {"type": "string"},
            "amount_usd": {"type": "number"}
          },
          "required": ["acquirer", "target", "amount_usd"],
          "additionalProperties": false
        }
      }
    }
  }'

Accuracy: no difference

Where there is one right answer, both got it. The water-tank problem (4.8 hours with all pipes open, then 12 more hours once pipe A is closed at half full), the Japanese factory-yield problem (1,610.4 good units, 13.16 hours, 11.96 hours) and the strawberry count (three) were all solved correctly by both.

Capabilities on FastMetal

Tencent added vision in Hy4, but images don't reach the model on our route, so it is text-only here.

Both bill per million input and output tokens, and Hy4 is the cheaper of the two. Current rates are on the Hy4 and Kimi K3.1 Pro model pages.

Which to use

• Latency matters, or a person reads the answer: Hy4. It comes back faster and explains its reasoning and alternatives.
• You need output in a fixed shape: Hy4 with a json_schema.
• You want only the deliverable, short, or you need to send images: Kimi K3.1 Pro.

The quickest way to decide is to read both answers to the same prompt on the comparison page.

FAQ

How do I make Hy4's answers shorter?

Tell it to return only the result, or pass a json_schema in response_format. You can also cut its reasoning with "reasoning": {"effort": "low"}.

Can Hy4 read images on FastMetal?

No. Use Kimi K3.1 Pro or another vision model from the catalog.

Did Hy4 really beat Kimi K3.1 Pro on finance?

That is Tencent's result, from an evaluation it ran with finance professionals and published on November 4, 2026. We have not measured finance tasks ourselves.

Are both available on the API?

Yes. Use hy4 or kimi-k3.1-pro as the model id with the OpenAI-compatible Chat Completions API, or pick either one in chat.