References
Disclaimer
Analysis on this site is based on publicly available information — research papers, open-source code, platform documentation, and published pricing — combined with empirical experiments run on the Tinker platform (see Empirical Validation). We have no inside knowledge of Tinker's internal architecture, actual infrastructure, or real cost structure. Our cost model, performance estimates, and comparisons are informed reconstructions calibrated against our own runs where possible.
The numbers may not reflect what Tinker (or any other platform) is actually doing under the hood. Treat everything here as an informed estimate, not ground truth. Where our experiments disagreed with the model (e.g. MFU was 1.6% not 35%), we updated the model — see the experiments page for calibration details. If you spot errors or have better data, we'd love to hear about it.
Tinker & Thinking Machines Lab
- Tinker Documentation — Thinking Machines Lab
- Tinker Cookbook — Thinking Machines Lab
- Tinker: Under the Hood — Thinking Machines Lab
- LoRA Without Regret — Thinking Machines Lab
- On-Policy Distillation (Tinker implementation) — Thinking Machines Lab
- Anatomy of a Finetuning API — Benjamin Anderson
Algorithms & training objectives
- Group Relative Policy Optimization (GRPO) — DeepSeekMath, Shao et al. (2024)
- DeepSeek-R1: Incentivizing Reasoning in LLMs via RL — DeepSeek (2025)
- It Takes Two: Your GRPO Is Secretly DPO — arXiv 2510.00977
- CISPO: Clipped Importance Sampling Policy Optimization — Chen et al. — arXiv 2506.13585
- CISPO (independent concurrent work) — Khatri et al. — arXiv 2510.13786
- Direct Reward Optimization — Richemond et al. — arXiv 2405.19107
- Kimi k1.5: Scaling Reinforcement Learning with LLMs — Kimi Team — arXiv 2501.12599
- Self-Distillation Policy Optimization — Gulcehre et al. — arXiv 2601.20802
- Learning to Discover at Test Time — arXiv 2601.16175
- On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes — Agarwal et al. — arXiv 2306.13649
Systems: async RL & pipeline parallelism
- PipelineRL: Faster On-policy RL for Long Sequence Generation — Piché et al., ServiceNow — arXiv 2509.19128
- VCPO: Variance Controlled Policy Optimization — arXiv 2602.17616
- Laminar: async RL framework — arXiv 2510.12633
- AReaL: async RL — arXiv 2505.24298
- RLHFuse — NSDI'25
- ROLLART: disaggregated rollout/training — arXiv 2512.22560
- DualPipe: bidirectional pipeline parallelism — DeepSeek
Inference & serving infrastructure
- Punica: SGMV multi-LoRA serving — MLSys'24
- S-LoRA: thousands of concurrent adapters — arXiv 2311.03285
- QeRL: NVFP4 quantization + LoRA — 2025
- DeepSeek V3 Technical Report — arXiv 2412.19437
Training runs & benchmarks
- SkyRL-Agent: Efficient RL Training for Multi-turn LLM Agent — arXiv 2511.16108
- DeepSWE: Training a Frontier Code Agent End-to-End with RL — Together AI (2025)
- INTELLECT-2: Globally Decentralized Reinforcement Learning — Prime Intellect — arXiv 2505.07291
- What went into training DeepSeek-R1? — Epoch AI
- The Era of Experience — Silver, Singh, Precup & Sutton — DeepMind (2025)
- SimpleRL-Zoo — 7B/14B RLVR in ~15 hrs on 16xH100
- DeepScaleR — 1.5B trained in 3,800 A100-hours, 43.1% AIME 2024
- Sebastian Raschka — "The State of Reinforcement Learning for LLM Reasoning"
RL frameworks & libraries
- veRL — Volcano Engine. Hybrid-controller model, FSDP/Megatron backend.
- OpenRLHF — Ray-based scheduling, DeepSpeed + vLLM.
- SLiME — Megatron-native with plugin system. Powers GLM-5/4.7.
- PRIME-RL — Decentralized async RL across 1000+ GPUs.
- TRL — HuggingFace. GRPOTrainer, PPOTrainer, DPO. Reference implementations.
- Unsloth — 2x faster, 70% less VRAM. Consumer GPU GRPO.
- torchforge — Meta/PyTorch-native RL post-training. Monarch actors, TorchStore, vLLM, TorchTitan. 512 H100s proven.
- SkyRL — Multi-turn agent RL with Harbor integration.
Benchmarks & environments
- OSWorld — GUI computer use benchmark, 369 tasks in full Ubuntu VMs (NeurIPS 2024)
- ComputerRL — RL training for computer use on OSWorld
- MATH — Mathematics problem solving dataset
- GSM8K — Grade school math word problems
- CodeContests — Competitive programming
- TerminalBench — Terminal-based agentic tasks
- Harbor — Sandbox lifecycle management for agentic RL
Pricing data
- ComputePrices.com — daily updated GPU pricing across 57 providers
- Cloud GPU pricing from Lambda, RunPod, Vast.ai, GCP, AWS, Azure
- Tinker pricing table (January 2026)