---
name: thalam-models
description: Pick, route, or call a THALAM model whenever the user asks for chat, coding, reasoning, long-context document work, vision, OCR, image generation, video generation, text-to-speech, embeddings, or reranking, or when the user explicitly mentions THALAM. Covers Claude, GPT, Gemini, DeepSeek, Qwen, Kimi, Llama, GLM, MiniMax, Seedance, Kling, ElevenLabs and more, exposed through a single OpenAI-compatible endpoint.
homepage: https://thalam.ai/docs/integrations/openclaw
---

# THALAM Models for OpenClaw

This skill connects OpenClaw to [THALAM](https://thalam.ai), a unified API gateway covering models from Anthropic, OpenAI, Google, Meta, Alibaba (Qwen), DeepSeek, MiniMax, Moonshot, Zhipu, ByteDance, Kuaishou, ElevenLabs, and more, through a single OpenAI-compatible endpoint. One API key. One bill. No lock-in.

## Setup

**1. Get a THALAM API key** at https://thalam.ai (sign-up takes under a minute, no credit card).

**2. Add THALAM as a model provider.** Open your config (default `~/.openclaw/openclaw.json` on Mac/Linux, `C:\Users\<you>\.openclaw\openclaw.json` on Windows — run `openclaw onboard` first if it doesn't exist) and merge the snippet below:

```json
{
  "models": {
    "mode": "merge",
    "providers": {
      "thalam": {
        "baseUrl": "https://api.thalam.ai/v1",
        "apiKey": "tl-live-...",
        "api": "openai-completions",
        "models": [
          {
            "id": "anthropic/claude-sonnet-4.6",
            "name": "Claude Sonnet 4.6 (THALAM)",
            "reasoning": false,
            "input": ["text", "image"],
            "cost": { "input": 0.003, "output": 0.015, "cacheRead": 0, "cacheWrite": 0 },
            "contextWindow": 200000,
            "maxTokens": 8192
          },
          {
            "id": "deepseek/deepseek-v4.1-flash",
            "name": "DeepSeek V4.1 Flash (THALAM)",
            "reasoning": false,
            "input": ["text", "image"],
            "cost": { "input": 0.0003, "output": 0.0012, "cacheRead": 0, "cacheWrite": 0 },
            "contextWindow": 1000000,
            "maxTokens": 8192
          }
        ]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": {
        "primary": "thalam/anthropic/claude-sonnet-4.6"
      },
      "models": {
        "thalam/anthropic/claude-sonnet-4.6": {},
        "thalam/deepseek/deepseek-v4.1-flash": {}
      }
    }
  }
}
```

**3. Validate and restart.** Run `openclaw config validate` to catch JSON errors, then `openclaw gateway restart` so the new provider takes effect. Reference any THALAM model in tasks using the prefixed id format `thalam/<provider>/<model-name>` (e.g., `thalam/deepseek/deepseek-v4.1-flash`).

**Important schema notes:**
- The provider object does **not** accept a top-level `name` field — display names live inside each `models` array entry.
- Every model id you want to call must appear in **both** `models.providers.thalam.models` (full definition) **and** `agents.defaults.models` (allowlist) — OpenClaw silently refuses to route to undeclared models.
- `mode: "merge"` is required so this provider config is added alongside any others (e.g. the default OpenAI plugin) instead of replacing them.

## Model selection guide

When choosing a model for a task, route by workload. Every id below is in the live catalog at https://thalam.ai/models.

### General reasoning / complex tasks
- `thalam/anthropic/claude-opus-4.8`: newest frontier-tier reasoning, long-form writing, agent orchestration. Use when correctness matters more than cost.
- `thalam/anthropic/claude-opus-4.7`: previous Opus generation, still available for version-pinned workloads.
- `thalam/openai/gpt-5.5`: OpenAI's flagship for general intelligence, coding and agentic tasks, with a 1M-token context.
- `thalam/openai/gpt-5.4`: strong tool use, vision input, broad multilingual including Arabic. 400K context.

### Cost-sensitive / high-volume
- `thalam/deepseek/deepseek-v4-flash-0731`: the official DeepSeek V4 Flash build at $0.14/1M input, with a 1M-token context. Use for classification, RAG and batch work.
- `thalam/qwen/qwen3.5-122b-a10b`: strong Arabic at a production price. Use for Arabic chat and batch translation.
- `thalam/openai/gpt-5-mini`: Western volume tier with 400K context.
- `thalam/google/gemma-4-31b-it`: Google's open model, dense and cheap.

### Reasoning (visible chain-of-thought)
- `thalam/deepseek/deepseek-r1-0528`: open-weight reasoning reference. Math, coding, step-by-step logic.
- `thalam/moonshotai/kimi-k2-thinking`: open-weight thinking model at a low price.
- `thalam/moonshotai/kimi-k3`: the strongest measured reasoning in the catalog, with a 1M-token context.

### Coding
- `thalam/anthropic/claude-sonnet-4.6`: default coding workhorse, strong tool use, long context, prompt caching.
- `thalam/deepseek/deepseek-v4-pro`: open-weight flagship with strong agentic coding, 1M context.
- `thalam/google/gemini-3.8-flash`: fast model built for software engineering and autonomous agents, 1M context.
- `thalam/openai/gpt-oss-120b`: open-weight 120B coding model at $0.10/1M input.

### Long context (input over 200K tokens)
- `thalam/google/gemini-2.5-pro`: 1M-token context, built for long-document analysis.
- `thalam/deepseek/deepseek-v4.1-flash`: 1M-token context and a 384K output ceiling, with image input.
- `thalam/openai/gpt-5.5`: 1M-token context at the frontier tier.

### Vision / multimodal input
- `thalam/zhipu/glm-4.6v`: vision-language model with strong document, OCR and chart understanding. It also reads short video clips.
- `thalam/openai/gpt-5.4`: frontier reasoning over images, 400K context.
- `thalam/deepseek/deepseek-v4.1-flash`: image input on a cheap 1M-context model.

### Image generation
- `thalam/google/gemini-3-pro-image`: Nano Banana Pro, frontier text-to-image with best-in-class in-image text for logos and posters.
- `thalam/openai/gpt-image-2.5-fast`: GPT Image 2.5 with five detail levels, from a cent-scale draft to a finished frame.
- `thalam/openai/gpt-image-2.5-precise`: edits up to four images and changes only what you ask for.
- `thalam/qwen/qwen-image-edit`: single-pass prompt edits at $0.02 per image.

### Video generation
- `thalam/bytedance/seedance-2.5`: 30-second single shots, up to 50 reference assets, native Arabic prompting and speech.
- `thalam/kuaishou/kling-v3.0-4k-t2v`: text-to-video in 4K.
- `thalam/minimax/hailuo-2.3-t2v`: current-generation Hailuo with sharper motion, $0.49 per video.

### Speech (text to speech)
- `thalam/elevenlabs/eleven-v3`: flagship expressive voice synthesis.
- `thalam/minimax/minimax-speech-2.8-hd`: top-tier multilingual speech with natural prosody.

### Embeddings / RAG
- `thalam/baai/bge-m3`: multilingual embedding, the RAG default, $0.01/1M.
- `thalam/qwen/qwen3-embedding-0.6b`: Arabic-strong embedding with a 32K window.
- `thalam/baai/bge-reranker-v2-m3`: multilingual reranker for the second stage of a RAG pipeline.

## Default routing logic

If the user does not specify a model, choose by task type:

1. **Simple Q&A or chat** → `thalam/deepseek/deepseek-v4-flash-0731` (cheapest capable)
2. **Complex reasoning, planning, multi-step analysis** → `thalam/anthropic/claude-opus-4.8`
3. **Code generation or refactor** → `thalam/anthropic/claude-sonnet-4.6`
4. **Documents over 200K tokens** → `thalam/google/gemini-2.5-pro`
5. **Image input / OCR / chart reading** → `thalam/zhipu/glm-4.6v`
6. **Image generation** → `thalam/google/gemini-3-pro-image`
7. **Video generation** → `thalam/bytedance/seedance-2.5`
8. **Text to speech** → `thalam/elevenlabs/eleven-v3`
9. **Embeddings for RAG** → `thalam/baai/bge-m3`

Always announce the chosen model at the start of the response so the user can override with a different one.

## Examples

**User:** "Summarise this 50-page PDF"
**Skill picks:** `thalam/google/gemini-2.5-pro` (1M-token context, built for long documents)

**User:** "Refactor this React component for accessibility"
**Skill picks:** `thalam/anthropic/claude-sonnet-4.6` (coding workhorse)

**User:** "Translate these 1000 product descriptions to Arabic"
**Skill picks:** `thalam/qwen/qwen3.5-122b-a10b` (strong Arabic at a production price)

**User:** "Generate a product video from this script"
**Skill picks:** `thalam/bytedance/seedance-2.5` (30-second shots with native speech)

## Pricing

Pay-per-token, no minimums, no lock-in. Full pricing at https://thalam.ai/pricing.

## Catalog updates

THALAM adds models regularly. The live catalog is at https://thalam.ai/models. If a model id in this skill is no longer available, fall back to the next entry in the same category.

## Support

- Docs: https://thalam.ai/docs
- Catalog: https://thalam.ai/models
- Status: https://thalam.ai/status
- Contact: support@thalam.ai