Making Inference Smarter
Pearl Research is building infrastructure and algorithms that deliver more tokens, faster, with AI economics built for scale, on any hardware - approaching the limit of intelligence-per-watt. Access leading open models through the most efficient inference engine.
- deepseek-ai/DeepSeek-V4-Flash-07311M
- deepseek-ai/DeepSeek-V4-Pro-08131M
- google/gemma-4-31B-it256k
- Qwen/Qwen3.8-27B256k
- Qwen/Qwen3.8-Flash-Next256k
- zai-org/GLM-5.31M
- zai-org/GLM-5.3-Flash1M
Pearl Research
Papers and engineering articles from the lab behind the serving stack.
- Hawkeye: Reproducing GPU-Level Non-DeterminismReplaying a GPU's exact arithmetic on a CPU, so any run can be audited against a deterministic reference.Read on arXiv
- A Fused Blockwise Hadamard Quantization Kernel in CuTe DSLSee how Pearl combines Hadamard transforms with integer quantization in a kernel built for production inference.Read on X
- 400x Faster Hashing on GPUsInside Pearl's blazingly fast BLAKE3 kernel that hashes GPU tensors 400x faster without a trip to the CPU.Read on X
Our Services
- Serverless
- OpenAI-compatible endpoints on our optimized serving stack. Pay per token, with no clusters to reserve, no idle GPUs, and nothing billing between requests.Get API Key
- Managed Inference
- Dedicated capacity for sustained workloads: reserved throughput, private routing, and custom or fine-tuned weights with better economics than the public endpoint.Book a Call
- On Premise
- Our serving stack deployed inside your own environment. Your hardware, your region, your controls, with inference traffic that never leaves your network.Book a Call
import osfrom openai import OpenAIclient = OpenAI(base_url="https://inference.pearlresearch.ai/v1",api_key=os.environ["PEARL_API_KEY"],)completion = client.chat.completions.create(model="deepseek-ai/DeepSeek-V4-Flash-0731",messages=[{"role": "user","content": "Explain proof of useful work.",}],)print(completion.choices[0].message.content)