TensorZero is an open-source stack for industrial-grade LLM applications:

• Gateway: access every LLM provider through a unified API, built for performance (<1ms p99 latency)
• Observability: store inferences and feedback in your database, available programmatically or in the UI
• Optimization: collect metrics and human feedback to optimize prompts, models, and inference strategies
• Evaluation: benchmark individual inferences or end-to-end workflows using heuristics, LLM judges, etc.
• Experimentation: ship with confidence with built-in A/B testing, routing, fallbacks, retries, etc.

Take what you need, adopt incrementally, and complement with other tools.

Website · Docs · Twitter · Slack · Discord

Quick Start (5min) · Deployment Guide · API Reference · Configuration Reference

Integrate with TensorZero once and access every major LLM provider.

• Call any LLM (API or self-hosted) through a single unified API
• Infer with streaming, tool use, structured outputs (JSON), batch, embeddings, multimodal (images, files), caching, etc.
• Create prompt templates and schemas to enforce a consistent, typed interface between your application and the LLMs
• Satisfy extreme throughput and latency needs, thanks to 🦀 Rust: <1ms p99 latency overhead at 10k+ QPS
• Use any programming language: integrate via our Python SDK, any OpenAI SDK, or our HTTP API
• Ensure high availability with routing, retries, fallbacks, load balancing, granular timeouts, etc.
• Enforce custom rate limits with granular scopes (e.g. user-defined tags) to keep usage under control
• Set up auth for TensorZero to allow clients to access models without sharing provider API keys
• Soon: spend tracking and budgeting

Supported Model Providers: Anthropic, AWS Bedrock, AWS SageMaker, Azure, DeepSeek, Fireworks, GCP Vertex AI Anthropic, GCP Vertex AI Gemini, Google AI Studio (Gemini API), Groq, Hyperbolic, Mistral, OpenAI, OpenRouter, SGLang, TGI, Together AI, vLLM, and xAI (Grok). Need something else? TensorZero also supports any OpenAI-compatible API (e.g. Ollama).

from tensorzero import TensorZeroGateway  # or AsyncTensorZeroGateway


with TensorZeroGateway.build_embedded(...) as t0:
    response = t0.inference(
        model_name="openai::gpt-4o-mini",
        # Try other providers easily: "anthropic::claude-sonnet-4-5"
        input={
            "messages": [
                {
                    "role": "user",
                    "content": "Write a haiku about TensorZero.",
                }
            ]
        },
    )

from openai import OpenAI  # or AsyncOpenAI
from tensorzero import patch_openai_client

client = OpenAI()

patch_openai_client(client, ...)

response = client.chat.completions.create(
    model="tensorzero::model_name::openai::gpt-4o-mini",
    # Try other providers easily: "tensorzero::model_name::anthropic::claude-sonnet-4-5"
    messages=[
        {
            "role": "user",
            "content": "Write a haiku about TensorZero.",
        }
    ],
)

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://localhost:3000/openai/v1",
});

const response = await client.chat.completions.create({
  model: "tensorzero::model_name::openai::gpt-4o-mini",
  // Try other providers easily: "tensorzero::model_name::anthropic::claude-sonnet-4-5"
  messages: [
    {
      role: "user",
      content: "Write a haiku about TensorZero.",
    },
  ],
});

curl -X POST "http://localhost:3000/inference" \
  -H "Content-Type: application/json" \
  -d '{
    "model_name": "openai::gpt-4o-mini",
    "input": {
      "messages": [
        {
          "role": "user",
          "content": "Write a haiku about TensorZero."
        }
      ]
    }
  }'

Zoom in to debug individual API calls, or zoom out to monitor metrics across models and prompts over time — all using the open-source TensorZero UI.

• Store inferences and feedback (metrics, human edits, etc.) in your own database
• Dive into individual inferences or high-level aggregate patterns using the TensorZero UI or programmatically
• Build datasets for optimization, evaluation, and other workflows
• Replay historical inferences with new prompts, models, inference strategies, etc.
• Export OpenTelemetry traces (OTLP) and export Prometheus metrics to your favorite application observability tools
• Soon: AI-assisted debugging and root cause analysis; AI-assisted data labeling

t0.experimental_list_inferences(
  function_name="sales_agent",
  filters=BooleanMetricFilter(
      metric_name="converted_sale",
      value=True,
  ),
  # + compound filters
  # + search
  # + pagination
  # ... and more ...
)

Send production metrics and human feedback to easily optimize your prompts, models, and inference strategies — using the UI or programmatically.

• Optimize your models with supervised fine-tuning, RLHF, and other techniques
• Optimize your prompts with automated prompt engineering algorithms like GEPA and MIPROv2
• Optimize your inference strategy with dynamic in-context learning, best/mixture-of-N sampling, etc.
• Enable a feedback loop for your LLMs: a data & learning flywheel turning production data into smarter, faster, and cheaper models
• Soon: synthetic data generation

Compare prompts, models, and inference strategies using evaluations powered by heuristics and LLM judges.

• Evaluate individual inferences with inference evaluations powered by heuristics or LLM judges (≈ unit tests for LLMs)
• Evaluate end-to-end workflows with workflow evaluations with complete flexibility (≈ integration tests for LLMs)
• Optimize LLM judges just like any other TensorZero function to align them to human preferences
• Soon: more built-in evaluators; headless evaluations

Evaluation » UI	Evaluation » CLI
	docker compose run --rm evaluations \ --evaluation-name extract_data \ --dataset-name hard_test_cases \ --variant-name gpt_4o \ --concurrency 5 Run ID: 01961de9-c8a4-7c60-ab8d-15491a9708e4 Number of datapoints: 100 ██████████████████████████████████████ 100/100 exact_match: 0.83 ± 0.03 (n=100) semantic_match: 0.98 ± 0.01 (n=100) item_count: 7.15 ± 0.39 (n=100)

Ship with confidence with built-in A/B testing, routing, fallbacks, retries, etc.

• Run adaptive A/B tests to ship with confidence and identify the best prompts and models for your use cases.
• Enforce principled experiments in complex workflows, including support for multi-turn LLM systems, sequential testing, and more.

Build with an open-source stack well-suited for prototypes but designed from the ground up to support the most complex LLM applications and deployments.

• Build simple applications or massive deployments with GitOps-friendly orchestration
• Extend TensorZero with built-in escape hatches, programmatic-first usage, direct database access, and more
• Integrate with third-party tools: specialized observability and evaluations, model providers, agent orchestration frameworks, etc.
• Iterate quickly by experimenting with prompts interactively using the Playground UI

How is TensorZero different from other LLM frameworks?

1. TensorZero enables you to optimize complex LLM applications based on production metrics and human feedback.
2. TensorZero supports the needs of industrial-grade LLM applications: low latency, high throughput, type safety, self-hosted, GitOps, customizability, etc.
3. TensorZero unifies the entire LLMOps stack, creating compounding benefits. For example, LLM evaluations can be used for fine-tuning models alongside AI judges.

Can I use TensorZero with ___?

Yes. Every major programming language is supported. It plays nicely with the OpenAI SDK, OpenTelemetry, and every major LLM.

Is TensorZero production-ready?

Yes. TensorZero is used by companies ranging from frontier AI startups to the Fortune 50.

Here's a case study: Automating Code Changelogs at a Large Bank with LLMs

How much does TensorZero cost?

TensorZero Stack (LLMOps platform) is 100% self-hosted and open-source.

TensorZero Autopilot (automated AI engineer) is a complementary paid product powered by the TensorZero Stack.

Who is building TensorZero?

Our technical team includes a former Rust compiler maintainer, machine learning researchers (Stanford, CMU, Oxford, Columbia) with thousands of citations, and the chief product officer of a decacorn startup. We're backed by the same investors as leading open-source projects (e.g. ClickHouse, CockroachDB) and AI labs (e.g. OpenAI, Anthropic). See our $7.3M seed round announcement and coverage from VentureBeat. We're hiring in NYC.

How do I get started?

You can adopt TensorZero incrementally. Our Quick Start goes from a vanilla OpenAI wrapper to a production-ready LLM application with observability and fine-tuning in just 5 minutes.

Watch LLMs get better at data extraction in real-time with TensorZero!

Dynamic in-context learning (DICL) is a powerful inference-time optimization available out of the box with TensorZero. It enhances LLM performance by automatically incorporating relevant historical examples into the prompt, without the need for model fine-tuning.

Start building today. The Quick Start shows it's easy to set up an LLM application with TensorZero.

Questions? Ask us on Slack or Discord.

Using TensorZero at work? Email us at [email protected] to set up a Slack or Teams channel with your team (free).

We are working on a series of complete runnable examples illustrating TensorZero's data & learning flywheel.

Optimizing Data Extraction (NER) with TensorZero

This example shows how to use TensorZero to optimize a data extraction pipeline. We demonstrate techniques like fine-tuning and dynamic in-context learning (DICL). In the end, an optimized GPT-4o Mini model outperforms GPT-4o on this task — at a fraction of the cost and latency — using a small amount of training data.

Agentic RAG — Multi-Hop Question Answering with LLMs

This example shows how to build a multi-hop retrieval agent using TensorZero. The agent iteratively searches Wikipedia to gather information, and decides when it has enough context to answer a complex question.

Writing Haikus to Satisfy a Judge with Hidden Preferences

This example fine-tunes GPT-4o Mini to generate haikus tailored to a specific taste. You'll see TensorZero's "data flywheel in a box" in action: better variants leads to better data, and better data leads to better variants. You'll see progress by fine-tuning the LLM multiple times.

Image Data Extraction — Multimodal (Vision) Fine-tuning

This example shows how to fine-tune multimodal models (VLMs) like GPT-4o to improve their performance on vision-language tasks. Specifically, we'll build a system that categorizes document images (screenshots of computer science research papers).

Improving LLM Chess Ability with Best-of-N Sampling

This example showcases how best-of-N sampling can significantly enhance an LLM's chess-playing abilities by selecting the most promising moves from multiple generated options.

We write about LLM engineering on the TensorZero Blog. Here are some of our favorite posts:

• Bandits in your LLM Gateway: Improve LLM Applications Faster with Adaptive Experimentation (A/B Testing)
• Is OpenAI's Reinforcement Fine-Tuning (RFT) Worth It?
• Distillation with Programmatic Data Curation: Smarter LLMs, 5-30x Cheaper Inference
• From NER to Agents: Does Automated Prompt Engineering Scale to Complex Tasks?