Tinker is a training API for researchers

Glean’s enterprise knowledge management relies on search. Frontier models are needed to reason about the results, but using them to gather the relevant documents is slow and expensive.

Glean chose NVIDIA Nemotron 3 Nano as the base model to minimize latency. Waldo was trained on Tinker in two steps: a DPO warmup on Glean’s anonymized production traces (sequences of preferred and rejected tool calls), followed by RL rewarded for recalling the documents actually cited in the final response and for using only the tools it needs.

Beyond accelerating search, Waldo gauges the reasoning effort a research task requires. Integrated into Glean’s full harness, it delivered 50% lower latency for search while using 25% fewer tokens, letting Glean capitalize on its unique training signal while leaving the infrastructure to Tinker.

Answering broad user questions about a corpus of documents requires accurate multi-hop reasoning. Relying on frontier models for multi-hop retrieval is costly, slow, and fills the context window with irrelevant or redundant information, taking valuable context away from the larger model’s higher-level goal. Chroma trained Context-1 as an efficient subagent for larger models in two stages:

1. SFT on quality-filtered agentic search trajectories from large models: reasoning steps, tool calls, and the final document set.
2. RL with a reward combining retrieval of the correct documents and token efficiency.

Context-1 was trained to manage its own context, pruning irrelevant documents as it searches, resulting in more efficient token trajectories. Tested on web search, email, finance filings, and legal documents, Context-1 matched or beat several frontier models at a fraction of the cost and up to 10x the tokens/sec. It generalized to domains outside its training, showing that a subagent trained for a task beats one a generalist spins up ad hoc.

Using Tinker, Lightning Rod trained a lightweight adapter on gpt-oss-120b, producing a specialized model for predicting clinical outcomes that runs on a single GPU. Calibration error dropped by roughly 70%, the Brier skill score rose from ~0% to 27%, and it won 84% of blind reasoning reviews against the base model. It also edged out GPT-5 on Brier score at a fraction of the size.

Mantic builds AI systems for judgmental forecasting — predicting geopolitical, economic, and technological events where there is no clean historical data to extrapolate from. After finding iteration on open-source RL frameworks too slow, they moved to Tinker and fine-tuned gpt-oss-120b with reinforcement learning on roughly 10,000 real-world forecasting questions.

The fine-tuned model went from trailing every frontier LLM to marginally outperforming them on the forecasting benchmark. Its predictions were decorrelated from those of the frontier models, making it the largest single contributor to the optimal ensemble.

The published algorithm assumes on-policy rollouts, and the naive off-policy fix, importance sampling, collapsed during training. Trajectory’s recipe, built on Tinker, adds clipped importance ratios and capped advantages and stays stable across hour-long trajectories where most rollouts fail. It lifted gpt-oss-120b from 5% to a 25% pass rate on the APEX-Agents benchmark, nine points beyond unaugmented SDPO, while training three steps stale for a 2x wall-clock speedup.

The obstacle to continual learning is catastrophic forgetting. Supervised fine-tuning is off-policy, so the model learns to imitate tokens it would never produce itself. As a result, every new skill learned this way erodes prior capabilities.

SDFT, introduced by Shenfeld, Damani, Hübotter et al., with support from a Tinker research grant, sidesteps the problem: place the expert demonstration in context, let the model generate its own on-policy version of the solution, and distill that behavior back into the demonstration-free model. Because the training signal stays on-policy, learning is faster than off-policy distillation and existing capabilities survive.

Across skill-learning benchmarks, SDFT beat standard SFT on every new task — 70.6% vs 63.2% on tool use — while substantially reducing forgetting, letting a single model accumulate skills sequentially where SFT regresses.

A chain of thought is only useful for oversight if it reflects the reasoning behind the answer. Models often rationalize instead: relying on a spurious cue while writing a plausible justification that never mentions it.

The Stanford team made faithfulness trainable: reward a chain of thought when a monitor that reads it can predict how the model will behave on counterfactual inputs (modified versions of the original prompts). With the training infrastructure abstracted away by Tinker, they iterated on rewards and evals across models up to 235B parameters, combining RL with LLM-guided rewriting of reasoning traces.

Monitors reading the fine-tuned models’ reasoning became 35 points more accurate at catching reliance on spurious cues.

Axiom Math builds AI for formal mathematics, where every claim has a checkable ground truth. That ground truth is what makes the domain trainable: Lean checks every step of a proof, and the reward is exact — a proof either verifies or it doesn’t. AxiomProver uses Tinker to train via RL and pursue that reward end to end, conjecturing, proving, and verifying its own work.

In December 2025 it solved all 12 problems on the Putnam, the world’s hardest college-level math competition: 8 by the end of the competition window and the remaining 4 in the days after, every proof formally verified in Lean. It came just days after AxiomProver generated verifiable solutions to two Erdős problems. The model now formalizes cutting-edge research in Lean, including recent papers in algebraic geometry.