tempera
Join waitlist

Fine-tuned LLMs for the enterprise

Post-train smaller, deployable models on the proprietary data and workflows frontier labs skip — financial, legal, ops, and beyond.

Fine-Tuning

Post-train on what frontier labs don't

Frontier RL environments don't cover your 10-Ks, your trading desks, or your underwriting playbooks. We do — turning your domain into a model that beats a vanilla commercial LLM on the work that matters to you.

R+0.73ppogrpodpo

Domain data

Proprietary corpora, traces, and tool calls become the training signal.

RSI

From reinforcement learning to real generalization

Our mission is to push post-training beyond reinforcement learning to true generalization — models that reason on your domain, not just on academic benchmarks. We get there by discovering and iterating state-of-the-art methods through recursive agentic self-improvement.

The loop

The post-training loop, automated

Three primitives, one closed loop, repeated until the model generalizes.

step 01→ next

Author

Define environments, actions, tools, and rewards in a typed, versioned API.

step 02→ next

Rollout

Run thousands of parallel rollouts; every step traced, every reward attributed.

step 03↺ repeat

Train

Tempera explores post-training methods on your data and ships the model that generalizes best.

API

Three calls. One loop

Compose environments, rollouts, and training in a single typed surface.

import { tempera } from "@tempera/sdk";

// Define the world your model trains in.
const env = await tempera.envs.create({
  name: "research-assistant",
  observation: { kind: "text" },
  action: {
    kind: "tool-use",
    tools: ["search", "shell", "browse"],
  },
  reward: { fn: "./rewards/helpfulness.ts" },
});

env.publish({ tag: "v1" });

Deployment

Your model. Your perimeter

Fine-tuned models ship where your data lives. Pick the isolation model that matches your security and compliance posture.

01

VPC · single-tenant

A dedicated control and data plane in your cloud account. No shared compute, no shared weights.

  • Runs in your AWS, GCP, or Azure VPC
  • Dedicated GPUs and inference endpoints
  • Data and weights never leave your account
02

VPC · multi-tenant

Shared managed control plane with isolated data planes per customer. Faster to onboard, lower TCO.

  • Tempera-managed control plane
  • Per-tenant isolated inference
  • SOC 2-aligned tenant boundaries
03

On-prem

Air-gapped deployment on your hardware for the most regulated environments.

  • Runs on your own GPU clusters
  • Offline / air-gapped supported
  • Bring your own KMS, IdP, and audit log sinks

Careers

Join us

Small team, large ideas, infinite mission.

Research

full-time · San Francisco in-person

Perform research on post-training, generalization, world models, and recursive self-improvement.

Apply → founders@tempera.dev

Research Infrastructure

full-time · San Francisco in-person

Build comprehensive scheduling, distributed training systems, and observability to power research.

Apply → founders@tempera.dev

Machine Learning

full-time · San Francisco in-person

Turn research into production.

Apply → founders@tempera.dev

Open call

full-time · San Francisco in-person

We're always looking for talented individuals across all disciplines to consider joining.

Apply → founders@tempera.dev

Waitlist

Bring your data. Ship a model

Join the waitlist for the first cohort, or email the founders directly.

Join the waitlistfounders@tempera.dev