OpenAI-Compatible APIs

OpenAI-Compatible APIs

ArtemisKit works with any provider that implements the OpenAI API specification. This includes cloud providers, local inference servers, and API aggregators.

How It Works

The @artemiskit/adapter-openai adapter accepts a baseUrl option that redirects API calls to any OpenAI-compatible endpoint:

providers:
  openai:
    apiKey: ${YOUR_API_KEY}
    baseUrl: https://your-provider.com/v1

Supported Providers

Provider	Base URL	Best For
Groq	https://api.groq.com/openai/v1	Ultra-fast LPU inference
Together AI	https://api.together.xyz/v1	200+ open-source models
Fireworks AI	https://api.fireworks.ai/inference/v1	Function calling, MoE models
OpenRouter	https://openrouter.ai/api/v1	400+ models, unified API
Ollama	http://localhost:11434/v1	Local inference
LM Studio	http://localhost:1234/v1	Local GUI + API
vLLM	http://localhost:8000/v1	Production self-hosting

---

Groq

Ultra-fast inference powered by custom LPU (Language Processing Unit) hardware. Achieves up to 1660 tokens/second with speculative decoding.

Setup

1. Get your API key from console.groq.com

2. Set environment variable:

export GROQ_API_KEY="gsk_..."

3. Configure ArtemisKit:

name: groq-test
provider: openai
model: llama-3.3-70b-specdec

providers:
  openai:
    apiKey: ${GROQ_API_KEY}
    baseUrl: https://api.groq.com/openai/v1

cases:
  - id: example
    prompt: "What is 2+2?"
    expected:
      type: contains
      values: ["4"]

Available Models

Model	Context	Notes
llama-3.3-70b-specdec	128K	Fastest 70B (1660 T/s with speculative decoding)
llama-3.3-70b-versatile	128K	Best quality general-purpose
llama-3-groq-70b-tool-use	8K	#1 on Berkeley Function Calling Leaderboard
llama-3.1-405b-instruct	128K	Largest open model
qwen3-32b	128K	Strong multilingual
mixtral-8x7b-32768	32K	Fast MoE

---

Together AI

Access to 200+ open-source models with the Together Inference Engine (4x faster than vLLM). Features frontier-class models including Qwen3 and DeepSeek.

Setup

1. Get your API key from api.together.xyz

2. Set environment variable:

export TOGETHER_API_KEY="..."

3. Configure ArtemisKit:

name: together-test
provider: openai
model: Qwen/Qwen3-235B-A22B-Instruct

providers:
  openai:
    apiKey: ${TOGETHER_API_KEY}
    baseUrl: https://api.together.xyz/v1

cases:
  - id: example
    prompt: "Explain quantum computing"
    expected:
      type: contains
      values: ["qubit"]

Available Models

Model	Context	Notes
Qwen/Qwen3-235B-A22B-Instruct	262K	Outperforms GPT-4 on benchmarks, 22B active params
Qwen/Qwen3-235B-A22B-Thinking	262K	Deep reasoning mode, beats OpenAI o3
Qwen/Qwen3-Coder-480B-A35B	128K	Largest open-source coding model
deepseek-ai/DeepSeek-R1	128K	State-of-the-art reasoning (math, code, logic)
deepseek-ai/DeepSeek-V3	128K	671B MoE, 37B active, tool calling
meta-llama/Llama-3.3-70B-Instruct-Turbo	128K	Fast Llama 3.3

---

Fireworks AI

Optimized for function calling, structured outputs, and MoE models. 4x throughput with 50% lower latency on NVIDIA H100s.

Setup

1. Get your API key from fireworks.ai/account/api-keys

2. Set environment variable:

export FIREWORKS_API_KEY="..."

3. Configure ArtemisKit:

name: fireworks-test
provider: openai
model: accounts/fireworks/models/qwen3-235b-a22b

providers:
  openai:
    apiKey: ${FIREWORKS_API_KEY}
    baseUrl: https://api.fireworks.ai/inference/v1

cases:
  - id: example
    prompt: "Write a haiku about testing"
    expected:
      type: contains
      values: ["test", "code"]
      mode: any

Available Models

Model	Context	Notes
accounts/fireworks/models/qwen3-235b-a22b	262K	Best quality MoE
accounts/fireworks/models/qwen3-coder-480b-a35b-instruct	128K	Largest coding model
accounts/fireworks/models/deepseek-r1-0528	128K	Advanced reasoning
accounts/fireworks/models/glm-4p7-flash	202K	30B MoE, fast and efficient
accounts/fireworks/models/llama-v3p3-70b-instruct	128K	Reliable general-purpose

---

OpenRouter

Single unified API for 400+ models from all major providers. Automatic fallbacks, load balancing, and cost optimization.

Setup

1. Get your API key from openrouter.ai/keys

2. Set environment variable:

export OPENROUTER_API_KEY="sk-or-..."

3. Configure ArtemisKit:

name: openrouter-test
provider: openai
model: deepseek/deepseek-r1

providers:
  openai:
    apiKey: ${OPENROUTER_API_KEY}
    baseUrl: https://openrouter.ai/api/v1

cases:
  - id: example
    prompt: "Hello"
    expected:
      type: contains
      values: ["hello", "hi"]
      mode: any

Available Models

OpenRouter provides access to models from all major providers:

Model	Provider	Notes
openai/gpt-5.2	OpenAI	400K context, adaptive reasoning
google/gemini-3	Google	1M token context, multimodal
anthropic/claude-sonnet-4	Anthropic	Latest Claude
deepseek/deepseek-r1	DeepSeek	Frontier quality at 1/100th cost
mistralai/devstral-2	Mistral	123B agentic coding, 256K context
meta-llama/llama-4-scout	Meta	Multimodal, extreme context
xiaomi/mimo-v2-flash	Xiaomi	309B MoE, 15B active, 256K context

See openrouter.ai/models for the full list of 400+ models.

---

Ollama

Run models locally on your machine. No API key required. Partners with OpenAI to bring GPT-OSS to local inference.

Setup

1. Install Ollama from ollama.com

2. Pull a model:

ollama pull qwen3

3. Start the server (runs automatically on install):

ollama serve

4. Configure ArtemisKit:

name: ollama-test
provider: openai
model: qwen3

providers:
  openai:
    apiKey: ollama  # Required but ignored
    baseUrl: http://localhost:11434/v1

cases:
  - id: example
    prompt: "What is the capital of France?"
    expected:
      type: contains
      values: ["Paris"]

Available Models

Run ollama list to see installed models. Popular options:

Model	Size	Notes
qwen3	8B	Latest Qwen with dual-mode reasoning
qwen3:72b	72B	Best quality Qwen
deepseek-r1	7B-671B	State-of-the-art reasoning
llama3.3	70B	Latest Llama
gpt-oss	varies	OpenAI’s open-source model
gemma3	4B-27B	Google’s efficient models
mistral-large-3	varies	Multimodal MoE for enterprise
qwen3-coder	varies	Code-specialized

---

LM Studio

Desktop app with a local API server. Easy model management with GUI.

Setup

1. Download from lmstudio.ai

2. Load a model in the app (recommended: Qwen3, DeepSeek-R1, or Llama 3.3)

3. Start the local server (Developer tab → Start Server)

4. Configure ArtemisKit:

name: lmstudio-test
provider: openai
model: local-model  # Model name from LM Studio

providers:
  openai:
    apiKey: lm-studio  # Required but ignored
    baseUrl: http://localhost:1234/v1

cases:
  - id: example
    prompt: "Explain recursion"
    expected:
      type: contains
      values: ["function", "itself"]
      mode: any

---

vLLM

High-performance inference server for production self-hosting. Supports speculative decoding, continuous batching, and PagedAttention.

Setup

1. Install vLLM:

pip install vllm

2. Start the server:

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen3-8B \
  --port 8000

3. Configure ArtemisKit:

name: vllm-test
provider: openai
model: Qwen/Qwen3-8B

providers:
  openai:
    apiKey: vllm  # Required but ignored
    baseUrl: http://localhost:8000/v1

cases:
  - id: example
    prompt: "Hello"
    expected:
      type: contains
      values: ["hello", "hi"]
      mode: any

---

Troubleshooting

Connection Refused

For local servers (Ollama, LM Studio, vLLM), ensure the server is running:

# Check if server is responding
curl http://localhost:11434/v1/models

Authentication Errors

Some providers require specific header formats. If you get auth errors, check the provider’s documentation for any custom headers.

Model Not Found

Model names vary by provider. Check the provider’s documentation for exact model identifiers.

Timeout Errors

For local inference with large models, increase the timeout:

providers:
  openai:
    baseUrl: http://localhost:11434/v1
    timeout: 120000  # 2 minutes

See Also

• OpenAI
• Azure OpenAI
• Scenario Format