peterindia.netLarge Reasoning Models
Models that spend extra compute thinking before they answer: proprietary and open-weight reasoning models, the methods used to train them, and the benchmarks that test them.
OpenAI's guide to its reasoning models, which think through a problem before answering and expose a reasoning-effort setting.
Best for: Hard coding, maths and planning tasks via API.
Visit OpenAI reasoning models →Anthropic's documentation for letting Claude reason step by step before responding, with a controllable thinking budget.
Best for: Complex analysis with a tunable thinking budget.
Visit Claude extended thinking →Gemini API documentation for the models' thinking capability, including thinking budgets and thought summaries.
Best for: Reasoning in the Gemini ecosystem.
Visit Google Gemini thinking →DeepSeek's open reasoning model trained with large-scale reinforcement learning, released with distilled smaller versions.
Best for: The reference open reasoning model.
Visit DeepSeek-R1 →Alibaba's 32-billion-parameter open reasoning model trained with reinforcement learning.
Best for: Capable reasoning at a manageable size.
Visit Qwen QwQ-32B →Moonshot AI's open-weight Kimi K2 family, including thinking models aimed at reasoning and agentic tasks.
Best for: Open models for reasoning plus tool use.
Visit Kimi K2 →Mistral AI's reasoning models, with an emphasis on multilingual step-by-step reasoning.
Best for: European and multilingual reasoning.
Visit Mistral Magistral →OpenAI's open-weight reasoning models, released for local and custom deployment.
Best for: Self-hosting a reasoning model with adjustable effort.
Visit OpenAI gpt-oss →Microsoft's small open reasoning model fine-tuned from Phi-4 for maths and science problems.
Best for: Reasoning on limited hardware.
Visit Phi-4-reasoning →NVIDIA's large open reasoning model derived from Llama, with a switchable reasoning mode.
Best for: Enterprise-scale reasoning on NVIDIA GPUs.
Visit Llama Nemotron Ultra →Ai2's fully open model family, including the Olmo 3 Think reasoning variants, released with data and training details.
Best for: Studying reasoning with open data and recipes.
Visit Olmo 3 →Hugging Face's open reproduction of the DeepSeek-R1 training pipeline.
Best for: Learning how reasoning models are trained.
Visit Open-R1 →Research code showing that a small reasoning dataset and a budget-forcing technique can improve test-time scaling.
Best for: Understanding test-time compute.
Visit s1: Simple test-time scaling →Competition and benchmark family built around ARC-AGI puzzles that test skill acquisition on novel tasks.
Best for: Testing generalisation beyond memorised patterns.
Visit ARC Prize →Multi-subject benchmark of expert-written questions designed to be hard for frontier models.
Best for: Measuring expert-level knowledge and reasoning.
Visit Humanity's Last Exam →Epoch AI's hub that tracks model results across benchmarks over time.
Best for: Comparing models across several benchmarks.
Visit Epoch AI Benchmarks →Benchmark and leaderboard that tests models on resolving real GitHub issues in software repositories.
Best for: Reasoning applied to code.
Visit SWE-bench →Match your requirement to a starting point, then compare vendors.
| Hard maths, code or planning | A proprietary reasoning model with a high effort or thinking budget |
|---|---|
| Control over data and cost | An open-weight reasoning model such as DeepSeek-R1, gpt-oss or Qwen QwQ |
| Reasoning on small hardware | A distilled or compact model such as Phi-4-reasoning |
| Quick conversational replies | A standard LLM, since reasoning models add latency and cost |
| Learn how they are trained | Open-R1 and the s1 research code |
| Judge claims | ARC Prize, Humanity's Last Exam and your own test cases |
Neighbouring directories and hubs.
Quick answers.
An LRM is a language model trained, often with reinforcement learning, to produce a chain of intermediate reasoning before its final answer, spending extra compute at inference time to solve harder problems.
A standard LLM answers directly from what it learned. An LRM deliberately works through steps, checks its work and can use a larger thinking budget, at the cost of more latency and tokens.
No. They help most on maths, code, logic and planning. For simple lookups, writing or chat they can be slower and costlier without being better.
Not necessarily. Visible reasoning text may not fully reflect how the model reached its answer, so verify conclusions independently.
It is the extra computation a model uses while answering, for example by thinking longer or sampling several attempts, rather than only during training.