llama-index-llms-thalam

LlamaIndex LLM integration for thalam. — one OpenAI-compatible endpoint for every leading AI model.

Built in the UAE. Text, image, and video models from a single endpoint, with per-key spend caps and audit logs as a built-in governance layer.

• One endpoint — DeepSeek, Qwen, Kimi, GLM, MiniMax, Claude, GPT, Llama, Grok, Mistral and more
• Drop-in LlamaIndex compatibility — RAG, query engines, agents, streaming, async all work unchanged
• Pay per token — no subscription, no minimums
• AED or USD invoicing on Enterprise contracts

Install

pip install llama-index-llms-thalam

Sign up at thalam.ai — no card required: new accounts get $1 in starter credit to use right away, then add a card to top up.

Quickstart

import os
from llama_index.llms.thalam import Thalam

os.environ["THALAM_API_KEY"] = "tl-..."

llm = Thalam(model="deepseek/deepseek-v3.2", temperature=0)

response = llm.complete("Explain quantum entanglement in 2 sentences.")
print(response.text)

Chat interface

from llama_index.core.llms import ChatMessage
from llama_index.llms.thalam import Thalam

llm = Thalam(model="deepseek/deepseek-v3.2")

response = llm.chat([
    ChatMessage(role="system", content="You are a senior MENA fintech analyst."),
    ChatMessage(role="user", content="Top 3 fintech trends in the UAE in 2026?"),
])
print(response.message.content)

Streaming

from llama_index.llms.thalam import Thalam

llm = Thalam(model="deepseek/deepseek-v3.2")

for chunk in llm.stream_complete("Write a haiku about Dubai."):
    print(chunk.delta, end="", flush=True)

Async

import asyncio
from llama_index.llms.thalam import Thalam

async def main():
    llm = Thalam(model="deepseek/deepseek-v3.2")
    response = await llm.acomplete("Hi")
    print(response.text)

asyncio.run(main())

Use as the global default LLM (RAG / query engines)

from llama_index.core import Settings, VectorStoreIndex, SimpleDirectoryReader
from llama_index.llms.thalam import Thalam

Settings.llm = Thalam(model="deepseek/deepseek-v3.2")

documents = SimpleDirectoryReader("./docs").load_data()
index = VectorStoreIndex.from_documents(documents)
query_engine = index.as_query_engine()

print(query_engine.query("Summarise these documents."))

Tool calling (function calling)

from llama_index.core.tools import FunctionTool
from llama_index.llms.thalam import Thalam

def get_weather(city: str) -> dict:
    """Get the weather in a city."""
    return {"temp": 32, "unit": "C", "city": city}

weather_tool = FunctionTool.from_defaults(fn=get_weather)

llm = Thalam(model="deepseek/deepseek-v4-pro")
response = llm.predict_and_call(
    [weather_tool],
    "What's the weather in Dubai?",
)
print(response)

Structured prediction

from pydantic import BaseModel, Field
from llama_index.llms.thalam import Thalam

class BlogMetadata(BaseModel):
    title: str = Field(description="The post title")
    tags: list[str] = Field(description="3-5 relevant tags")

llm = Thalam(model="deepseek/deepseek-v3.2")
result = llm.structured_predict(
    BlogMetadata,
    "Generate metadata for a blog post about MENA fintech.",
)
print(result.title, result.tags)

Available models

The full live catalog is at thalam.ai/models or via:

import os, requests
r = requests.get(
    "https://api.thalam.ai/v1/models",
    headers={"Authorization": f"Bearer {os.environ['THALAM_API_KEY']}"},
)
for m in r.json()["data"]:
    print(m["id"], "-", m["display_name"], "-", m["category"])

Common text-model IDs:

Configuration

from llama_index.llms.thalam import Thalam

llm = Thalam(
    model="deepseek/deepseek-v3.2",
    api_key="tl-...",                              # default: env THALAM_API_KEY
    api_base="https://api.thalam.ai/v1",            # default
    context_window=131072,                          # set 1_000_000 for v4-pro
    is_function_calling_model=True,
    temperature=0.7,
    max_tokens=512,
    timeout=60,
    max_retries=2,
)

Every keyword argument that llama_index.llms.openai_like.OpenAILike accepts is supported.

Context window note

The default context_window is conservative at 128k. For 1M-context models like deepseek/deepseek-v4-pro, pass context_window=1_000_000 to let LlamaIndex pack the full prompt budget.

Links

• thalam.ai
• Documentation
• Model catalog
• LlamaIndex docs

License

Apache-2.0