> ## Documentation Index
> Fetch the complete documentation index at: https://docs.kymaapi.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Which model should I use?

> Choose the right Kyma model by task, constraint, and tradeoff.

## Start here

If you do not want to think too hard about model choice:

* **Start with [`qwen-3.6-plus`](/models/qwen-3.6-plus)** for the best default across general work, coding, and reasoning.
* **Switch to [`kimi-k2.6`](/models/kimi-k2.6)** if your workflow looks like an agent with tools, long sessions, or screenshots.
* **Switch to [`deepseek-v4-pro`](/models/deepseek-v4-pro)** if you need top reasoning with 1M context.
* **Switch to [`deepseek-v4-flash`](/models/deepseek-v4-flash)** if you want V4-tier behavior at the lowest price.
* **Switch to [`gemini-2.5-flash`](/models/gemini-2.5-flash)** if context length is the main problem.

## Quick picks

| If you need...                                        | First pick                                                        | Why                                                                   | Second pick                                                               |
| ----------------------------------------------------- | ----------------------------------------------------------------- | --------------------------------------------------------------------- | ------------------------------------------------------------------------- |
| One default model                                     | [`qwen-3.6-plus`](/models/qwen-3.6-plus)                          | Best overall quality and safest default                               | [`deepseek-v4-flash`](/models/deepseek-v4-flash)                          |
| Newest flagship (vision + 1M ctx)                     | [`qwen-3.7-plus`](/models/qwen-3.7-plus)                          | Successor to qwen-3.6-plus — adds vision, 1M context, cheaper         | [`qwen-3.6-plus`](/models/qwen-3.6-plus)                                  |
| Best value                                            | [`deepseek-v4-flash`](/models/deepseek-v4-flash)                  | V4-tier quality, 1M context, native reasoning, lowest V4 price        | [`deepseek-v3`](/models/deepseek-v3)                                      |
| Top reasoning                                         | [`deepseek-v4-pro`](/models/deepseek-v4-pro)                      | 1.6T MoE flagship, 1M context, native reasoning                       | [`deepseek-r1`](/models/deepseek-r1)                                      |
| Deep reasoning                                        | [`deepseek-r1`](/models/deepseek-r1)                              | Best for logic, math, hard analysis                                   | [`deepseek-v4-pro`](/models/deepseek-v4-pro)                              |
| Tool-heavy agents                                     | [`kimi-k2.6`](/models/kimi-k2.6)                                  | Strong tool use, long context, multimodal                             | [`glm-5.1`](/models/glm-5.1)                                              |
| Long-running coding agents                            | [`glm-5.1`](/models/glm-5.1)                                      | Better for repo-scale engineering and sustained execution             | [`minimax-m2.5`](/models/minimax-m2.5)                                    |
| Agentic coding                                        | [`minimax-m2.5`](/models/minimax-m2.5)                            | Strong engineering workflow fit for typical coding agents             | [`glm-5.1`](/models/glm-5.1)                                              |
| Agentic coding, 1M context                            | [`minimax-m3`](/models/minimax-m3)                                | Newest MiniMax — multimodal input, sparse attention for long sessions | [`minimax-m2.5`](/models/minimax-m2.5)                                    |
| Massive open-weight reasoning                         | [`nemotron-3-ultra-550b`](/models/nemotron-3-ultra-550b)          | Largest US open-weight (550B MoE), 1M context, fast                   | [`deepseek-v4-pro`](/models/deepseek-v4-pro)                              |
| Fast coding loops                                     | [`qwen-3-32b`](/models/qwen-3-32b)                                | Lower latency while staying strong on code                            | [`qwen-3-coder`](/models/qwen-3-coder)                                    |
| 1M context                                            | [`gemini-2.5-flash`](/models/gemini-2.5-flash)                    | Cheapest long-context option                                          | [`gemini-3-flash`](/models/gemini-3-flash)                                |
| Vision / screenshots                                  | [`gemma-4-31b`](/models/gemma-4-31b)                              | Cheapest solid multimodal option                                      | [`kimi-k2.6`](/models/kimi-k2.6)                                          |
| Live web search / current info                        | [`sonar`](/models/sonar)                                          | Real-time web search with citations, no RAG to build                  | [`sonar-pro`](/models/sonar-pro)                                          |
| Cheap bulk automation                                 | [`glm-4.5-air`](/models/glm-4.5-air)                              | Low-cost agentic path                                                 | [`glm-4.7-flash`](/models/glm-4.7-flash)                                  |
| Cheap long-context throughput                         | [`glm-4.7-flash`](/models/glm-4.7-flash)                          | Fast, efficient, long context                                         | [`gemini-2.5-flash`](/models/gemini-2.5-flash)                            |
| Cheap multimodal throughput                           | [`step-3.7-flash`](/models/step-3.7-flash)                        | Text+image+video input, tool calling, 256K context                    | [`gemma-4-31b`](/models/gemma-4-31b)                                      |
| Default image generation                              | [`recraft-v4`](/models/image-generation#recraft-v4)               | #1 HF Arena, design-quality, \$0.054                                  | [`flux-2-pro`](/models/image-generation#flux-2-pro)                       |
| Photoreal / hero shots                                | [`flux-2-pro`](/models/image-generation#flux-2-pro)               | BFL 32B, multi-reference, gen+edit                                    | [`recraft-v4-pro`](/models/image-generation#recraft-v4-pro)               |
| Multi-reference blend                                 | [`flux-2-pro`](/models/image-generation#flux-2-pro)               | Up to 10 source images via `image_urls`                               | —                                                                         |
| Image edit / inpaint                                  | [`flux-kontext-pro`](/models/image-generation#flux-kontext-pro)   | Image-to-image editor, requires `image_url`                           | [`flux-2-pro`](/models/image-generation#flux-2-pro)                       |
| Logos, typography, packaging                          | [`ideogram-v3`](/models/image-generation#ideogram-v3)             | Best legible-text image model                                         | [`recraft-v4-vector`](/models/image-generation#recraft-v4-vector)         |
| Native SVG / vector                                   | [`recraft-v4-vector`](/models/image-generation#recraft-v4-vector) | True paths + layers, edit in Figma                                    | [`recraft-v4-vector-pro`](/models/image-generation#recraft-v4-vector-pro) |
| Print-ready design (4MP)                              | [`recraft-v4-pro`](/models/image-generation#recraft-v4-pro)       | V4 quality at 4MP for print                                           | [`recraft-v4-vector-pro`](/models/image-generation#recraft-v4-vector-pro) |
| Expressive TTS (character voices, emotional dialogue) | [`eleven-v3`](/api-reference/audio-speech)                        | Audio tags, emotional range, 70+ languages                            | [`eleven-multilingual-v2`](/api-reference/audio-speech)                   |
| Hero-quality TTS (narration)                          | [`eleven-multilingual-v2`](/api-reference/audio-speech)           | 29 languages, expressive, brand-safe                                  | [`eleven-turbo-v2-5`](/api-reference/audio-speech)                        |
| Real-time voice agent                                 | [`eleven-flash-v2-5`](/api-reference/audio-speech)                | \~75ms time-to-first-byte, 32 lang                                    | [`eleven-turbo-v2-5`](/api-reference/audio-speech)                        |
| Music generation                                      | [`elevenlabs-music`](/api-reference/audio-music)                  | Prompt-driven, lyrics support, 1s..5min                               | —                                                                         |
| Sound effects                                         | [`elevenlabs-sfx`](/api-reference/audio-sfx)                      | Whoosh, explosion, ambient — 0.5..22s                                 | —                                                                         |

<Tip>
  These are decision shortcuts, not absolute rankings. If your workload changes, your best model changes too.
</Tip>

## Choose by constraint

<AccordionGroup>
  <Accordion title="I want the safest default" icon="trophy">
    Pick **[`qwen-3.6-plus`](/models/qwen-3.6-plus)**.

    Use it when you want one model that is strong at general work, coding, reasoning, and multilingual tasks without forcing a lot of tradeoff thinking.
  </Accordion>

  <Accordion title="I care most about value" icon="coins">
    Pick **[`deepseek-v4-flash`](/models/deepseek-v4-flash)**.

    It is the best first stop when you want strong quality at lower cost — 1M context, native reasoning, MIT license. If you need an even cheaper lane for routine workloads, look at [`gpt-oss-120b`](/models/gpt-oss-120b) or [`glm-4.5-air`](/models/glm-4.5-air). If you have a stable production workload already on [`deepseek-v3`](/models/deepseek-v3), it stays available.
  </Accordion>

  <Accordion title="I need top reasoning at flagship quality" icon="brain-circuit">
    Pick **[`deepseek-v4-pro`](/models/deepseek-v4-pro)**.

    It is the V4 flagship — 1.6T MoE, 1M context, native reasoning. Best fit for complex coding, multi-step analysis, and research-grade work where quality wins over latency.
  </Accordion>

  <Accordion title="I need deep reasoning" icon="brain">
    Pick **[`deepseek-r1`](/models/deepseek-r1)**.

    Use it for difficult analysis, logic, math, and multi-step planning where you want a slower, deeper chain-of-thought trace. If it feels too slow, fall back to [`deepseek-v4-pro`](/models/deepseek-v4-pro) or [`qwen-3.6-plus`](/models/qwen-3.6-plus).
  </Accordion>

  <Accordion title="I need the best model for an agent" icon="robot">
    Start with **[`kimi-k2.6`](/models/kimi-k2.6)**.

    It is the best first pick when your workload uses tools, screenshots, or long multi-step sessions. If you want a text-only engineering alternative for repo-scale work, try [`glm-5.1`](/models/glm-5.1).
  </Accordion>

  <Accordion title="I need a long-running coding agent" icon="microchip">
    Start with **[`glm-5.1`](/models/glm-5.1)**.

    It is the better fit when the work is repo-scale, multi-file, and long-horizon. If your agent is more typical day-to-day coding than sustained engineering execution, fall back to [`minimax-m2.5`](/models/minimax-m2.5).
  </Accordion>

  <Accordion title="I need agentic coding specifically" icon="terminal-square">
    Start with **[`minimax-m2.5`](/models/minimax-m2.5)**.

    It fits engineering workflows well for normal coding-agent usage. If you want a newer productivity-oriented variant, try [`minimax-m2.7`](/models/minimax-m2.7). If you want a stronger long-horizon engineering model, move up to [`glm-5.1`](/models/glm-5.1).
  </Accordion>

  <Accordion title="I need a fast coding model" icon="bolt">
    Pick **[`qwen-3-32b`](/models/qwen-3-32b)**.

    It is the best fit for tight edit-run-debug loops. If you want a more code-specialized model with more context, try [`qwen-3-coder`](/models/qwen-3-coder).
  </Accordion>

  <Accordion title="I need long context" icon="file-lines">
    Pick **[`gemini-2.5-flash`](/models/gemini-2.5-flash)** for the safest long-context default.

    If you want a newer preview path, use [`gemini-3-flash`](/models/gemini-3-flash). If you want cheaper long-context throughput without needing multimodal input, look at [`glm-4.7-flash`](/models/glm-4.7-flash).
  </Accordion>

  <Accordion title="I need vision or screenshot understanding" icon="image">
    Start with **[`gemma-4-31b`](/models/gemma-4-31b)**.

    It is the cheapest strong multimodal option. If you also need stronger agent behavior, upgrade to [`kimi-k2.6`](/models/kimi-k2.6).
  </Accordion>

  <Accordion title="I need the cheapest route for bulk work" icon="wallet">
    Start with **[`glm-4.5-air`](/models/glm-4.5-air)**.

    It is the cheaper agentic lane for repeated automation tasks. If the workload is more about long context and throughput than agent behavior, try [`glm-4.7-flash`](/models/glm-4.7-flash).
  </Accordion>

  <Accordion title="I need live, cited web answers" icon="globe">
    Start with **[`sonar`](/models/sonar)** (alias `search`).

    It runs a real-time web search on every request and returns a current answer with citations — no search + scrape + RAG pipeline to build. For complex, multi-step research and longer cited reports, move up to [`sonar-pro`](/models/sonar-pro). Note both bill a small flat web-search fee on top of tokens, and neither supports tool calling — the search happens internally.
  </Accordion>
</AccordionGroup>

## Tradeoffs that matter

| Model                                            | Main strength                                         | Main tradeoff                                                      |
| ------------------------------------------------ | ----------------------------------------------------- | ------------------------------------------------------------------ |
| [`qwen-3.6-plus`](/models/qwen-3.6-plus)         | Best overall default                                  | Not the cheapest or fastest                                        |
| [`deepseek-v4-pro`](/models/deepseek-v4-pro)     | Top reasoning, 1M context, native reasoning           | Premium price, preview stage                                       |
| [`deepseek-v4-flash`](/models/deepseek-v4-flash) | Best value V4, 1M context, native reasoning           | Preview stage                                                      |
| [`deepseek-v3`](/models/deepseek-v3)             | Stable previous-gen flagship                          | Smaller context, no native reasoning                               |
| [`deepseek-r1`](/models/deepseek-r1)             | Best chain-of-thought reasoning                       | Slower                                                             |
| [`kimi-k2.6`](/models/kimi-k2.6)                 | Best tool-heavy agent behavior                        | Premium cost                                                       |
| [`minimax-m2.5`](/models/minimax-m2.5)           | Strong engineering workflows                          | Less general-purpose than Qwen flagship                            |
| [`qwen-3-32b`](/models/qwen-3-32b)               | Fast coding                                           | Shorter context                                                    |
| [`gemini-2.5-flash`](/models/gemini-2.5-flash)   | 1M context at good price                              | Not the strongest default for coding agents                        |
| [`gemma-4-31b`](/models/gemma-4-31b)             | Cheap vision                                          | Weaker than flagship text models on hard reasoning                 |
| [`glm-5.1`](/models/glm-5.1)                     | Long-running coding agents and repo-scale engineering | Premium lane, text-only, less battle-tested in Kyma than Qwen/Kimi |
| [`glm-4.5-air`](/models/glm-4.5-air)             | Cheap agentic bulk tasks                              | Lower ceiling than flagship models                                 |
| [`glm-4.7-flash`](/models/glm-4.7-flash)         | Cheap long-context throughput                         | Preview-stage model                                                |

## Use by task

| Task                                | First pick                                     | Fallback                                         |
| ----------------------------------- | ---------------------------------------------- | ------------------------------------------------ |
| General chat / assistant            | [`qwen-3.6-plus`](/models/qwen-3.6-plus)       | [`deepseek-v4-flash`](/models/deepseek-v4-flash) |
| Coding assistant                    | [`qwen-3.6-plus`](/models/qwen-3.6-plus)       | [`qwen-3-32b`](/models/qwen-3-32b)               |
| Autonomous coding agent             | [`kimi-k2.6`](/models/kimi-k2.6)               | [`minimax-m2.5`](/models/minimax-m2.5)           |
| Repo-scale engineering work         | [`glm-5.1`](/models/glm-5.1)                   | [`deepseek-v4-pro`](/models/deepseek-v4-pro)     |
| Math / reasoning / hard analysis    | [`deepseek-v4-pro`](/models/deepseek-v4-pro)   | [`deepseek-r1`](/models/deepseek-r1)             |
| Long document summarization         | [`gemini-2.5-flash`](/models/gemini-2.5-flash) | [`glm-4.7-flash`](/models/glm-4.7-flash)         |
| Data extraction / structured output | [`qwen-3-32b`](/models/qwen-3-32b)             | [`glm-4.5-air`](/models/glm-4.5-air)             |
| Screenshot / image understanding    | [`gemma-4-31b`](/models/gemma-4-31b)           | [`kimi-k2.6`](/models/kimi-k2.6)                 |
| Live web search / current info      | [`sonar`](/models/sonar)                       | [`sonar-pro`](/models/sonar-pro)                 |
| Cheap automation                    | [`glm-4.5-air`](/models/glm-4.5-air)           | [`gpt-oss-120b`](/models/gpt-oss-120b)           |

## Still not sure?

* Use alias `best` for `qwen-3.6-plus`
* Use alias `agent` for `kimi-k2.6`
* Use alias `reasoning` for `deepseek-r1`
* Use alias `long-context` for `gemini-2.5-flash`

See [model aliases](/guides/model-aliases) and [all models](/models/overview) for the canonical live catalog.