Skip to content
Python

feat(llm): add NVIDIA NIM, Kimi, Groq, and Mistral providers

Yijia-Xiao295e84cd546a
diff --git a/.env.example b/.env.example
index 8f321e548..aad73c9ad 100644
--- a/.env.example
+++ b/.env.example
@@ -11,6 +11,15 @@ ZHIPU_CN_API_KEY=
MINIMAX_API_KEY=
MINIMAX_CN_API_KEY=
OPENROUTER_API_KEY=
+MISTRAL_API_KEY=
+MOONSHOT_API_KEY=
+GROQ_API_KEY=
+NVIDIA_API_KEY=
+
+# Optional: a custom OpenAI-compatible endpoint (vLLM, LM Studio, llama.cpp,
+# relay). Select provider "openai_compatible" and set the base URL; the key is
+# optional (local servers need none).
+#OPENAI_COMPATIBLE_API_KEY=
# Optional: point at a remote Ollama server. When unset, defaults to
# the local instance at http://localhost:11434/v1. Convention follows
diff --git a/README.md b/README.md
index a84f1027f..20147d101 100644
--- a/README.md
+++ b/README.md
@@ -225,7 +225,7 @@ from tradingagents.graph.trading_graph import TradingAgentsGraph
from tradingagents.default_config import DEFAULT_CONFIG
config = DEFAULT_CONFIG.copy()
-config["llm_provider"] = "openai" # openai, google, anthropic, xai, deepseek, qwen, qwen-cn, glm, glm-cn, minimax, minimax-cn, openrouter, ollama, azure
+config["llm_provider"] = "openai" # e.g. openai, google, anthropic, deepseek, groq, ollama; openai_compatible covers any OpenAI-compatible endpoint (vLLM, LM Studio, llama.cpp, ...)
config["deep_think_llm"] = "gpt-5.5" # Model for complex reasoning
config["quick_think_llm"] = "gpt-5.4-mini" # Model for quick tasks
config["max_debate_rounds"] = 2
diff --git a/cli/utils.py b/cli/utils.py
index 6054956ce..3ee2d25c7 100644
--- a/cli/utils.py
+++ b/cli/utils.py
@@ -311,6 +311,10 @@ def _llm_provider_table() -> list[tuple[str, str, str | None]]:
("GLM", "glm", "https://open.bigmodel.cn/api/paas/v4/"),
("MiniMax", "minimax", "https://api.minimax.io/v1"),
("OpenRouter", "openrouter", "https://openrouter.ai/api/v1"),
+ ("Mistral", "mistral", "https://api.mistral.ai/v1"),
+ ("Kimi (Moonshot)", "kimi", "https://api.moonshot.ai/v1"),
+ ("Groq", "groq", "https://api.groq.com/openai/v1"),
+ ("NVIDIA NIM", "nvidia", "https://integrate.api.nvidia.com/v1"),
("Azure OpenAI", "azure", None),
("Ollama", "ollama", ollama_url),
("OpenAI-compatible (vLLM, LM Studio, llama.cpp, custom relay)", "openai_compatible", None),
diff --git a/tests/test_provider_registry.py b/tests/test_provider_registry.py
index 23f36c060..596108a77 100644
--- a/tests/test_provider_registry.py
+++ b/tests/test_provider_registry.py
@@ -35,6 +35,10 @@ def test_registry_membership():
("minimax", "https://api.minimax.io/v1", MinimaxChatOpenAI, False),
("minimax-cn", "https://api.minimaxi.com/v1", MinimaxChatOpenAI, False),
("openrouter", "https://openrouter.ai/api/v1", NormalizedChatOpenAI, False),
+ ("mistral", "https://api.mistral.ai/v1", NormalizedChatOpenAI, False),
+ ("kimi", "https://api.moonshot.ai/v1", NormalizedChatOpenAI, False),
+ ("groq", "https://api.groq.com/openai/v1", NormalizedChatOpenAI, False),
+ ("nvidia", "https://integrate.api.nvidia.com/v1", NormalizedChatOpenAI, False),
("ollama", "http://localhost:11434/v1", NormalizedChatOpenAI, False),
])
def test_registry_spec(provider, base_url, chat_class, responses):
diff --git a/tradingagents/llm_clients/api_key_env.py b/tradingagents/llm_clients/api_key_env.py
index c9353963a..4d909c400 100644
--- a/tradingagents/llm_clients/api_key_env.py
+++ b/tradingagents/llm_clients/api_key_env.py
@@ -30,6 +30,12 @@ PROVIDER_API_KEY_ENV: dict[str, Optional[str]] = {
"minimax": "MINIMAX_API_KEY",
"minimax-cn": "MINIMAX_CN_API_KEY",
"openrouter": "OPENROUTER_API_KEY",
+ # Additional hosted OpenAI-compatible providers (model is user-specified).
+ # kimi -> Moonshot AI; nvidia -> NVIDIA NIM.
+ "mistral": "MISTRAL_API_KEY",
+ "kimi": "MOONSHOT_API_KEY",
+ "groq": "GROQ_API_KEY",
+ "nvidia": "NVIDIA_API_KEY",
# Local runtimes do not authenticate.
"ollama": None,
# Generic OpenAI-compatible endpoint: the client reads this when set (keyed
diff --git a/tradingagents/llm_clients/model_catalog.py b/tradingagents/llm_clients/model_catalog.py
index e25e2feeb..f1fc45b09 100644
--- a/tradingagents/llm_clients/model_catalog.py
+++ b/tradingagents/llm_clients/model_catalog.py
@@ -7,6 +7,13 @@ from typing import Dict, List, Tuple
ModelOption = Tuple[str, str]
ProviderModeOptions = Dict[str, Dict[str, List[ModelOption]]]
+# Providers that serve many / frequently-changing models: offer only "Custom
+# model ID" rather than a list that goes stale.
+_CUSTOM_ONLY: Dict[str, List[ModelOption]] = {
+ "quick": [("Custom model ID", "custom")],
+ "deep": [("Custom model ID", "custom")],
+}
+
# Shared model list for GLM via Z.AI (international) and BigModel (China).
# Source: docs.z.ai (GLM Coding Plan supported models + LLM guides).
@@ -177,10 +184,15 @@ MODEL_OPTIONS: ProviderModeOptions = {
},
# Generic OpenAI-compatible endpoint: the model is whatever the user's
# server serves, so only "Custom model ID" is offered.
- "openai_compatible": {
- "quick": [("Custom model ID", "custom")],
- "deep": [("Custom model ID", "custom")],
- },
+ "openai_compatible": _CUSTOM_ONLY,
+ # Hosted OpenAI-compatible providers that serve many (and frequently
+ # changing) models — offer "Custom model ID" rather than a list that goes
+ # stale. The endpoint + key are wired by the provider; the user picks the
+ # model their account has access to.
+ "mistral": _CUSTOM_ONLY,
+ "kimi": _CUSTOM_ONLY,
+ "groq": _CUSTOM_ONLY,
+ "nvidia": _CUSTOM_ONLY,
}
diff --git a/tradingagents/llm_clients/openai_client.py b/tradingagents/llm_clients/openai_client.py
index 262a6e55a..8aaa988df 100644
--- a/tradingagents/llm_clients/openai_client.py
+++ b/tradingagents/llm_clients/openai_client.py
@@ -189,6 +189,10 @@ OPENAI_COMPATIBLE_PROVIDERS: dict[str, ProviderSpec] = {
"minimax": ProviderSpec(base_url="https://api.minimax.io/v1", chat_class=MinimaxChatOpenAI),
"minimax-cn": ProviderSpec(base_url="https://api.minimaxi.com/v1", chat_class=MinimaxChatOpenAI),
"openrouter": ProviderSpec(base_url="https://openrouter.ai/api/v1"),
+ "mistral": ProviderSpec(base_url="https://api.mistral.ai/v1"),
+ "kimi": ProviderSpec(base_url="https://api.moonshot.ai/v1"),
+ "groq": ProviderSpec(base_url="https://api.groq.com/openai/v1"),
+ "nvidia": ProviderSpec(base_url="https://integrate.api.nvidia.com/v1"),
"ollama": ProviderSpec(base_url="http://localhost:11434/v1", base_url_env="OLLAMA_BASE_URL",
key_optional=True, placeholder_key="ollama"),
# Generic endpoint: user supplies base_url; key optional (keyless local).
diff --git a/tradingagents/llm_clients/validators.py b/tradingagents/llm_clients/validators.py
index 8d04a9384..bed09a51a 100644
--- a/tradingagents/llm_clients/validators.py
+++ b/tradingagents/llm_clients/validators.py
@@ -3,9 +3,13 @@
from .model_catalog import get_known_models
-# Providers whose model names are user-defined (local servers, relays, any
-# OpenAI-compatible endpoint), so any model string is accepted without warning.
-_ANY_MODEL_PROVIDERS = ("ollama", "openrouter", "openai_compatible")
+# Providers whose model names are user-defined (local servers, relays, hosted
+# OpenAI-compatible endpoints serving many models), so any model string is
+# accepted without warning.
+_ANY_MODEL_PROVIDERS = (
+ "ollama", "openrouter", "openai_compatible",
+ "mistral", "kimi", "groq", "nvidia",
+)
VALID_MODELS = {
provider: models