Reasoning LLMs and LRMs logopeterindia.netReasoning LLMs & LRMs
Home / Generative AI / Reasoning LLMs and Large Reasoning Models (LRMs)
Generative AI ยท 2026

Reasoning LLMs and Large Reasoning Models (LRMs)

Language models that spend extra compute thinking before they answer: proprietary and open-weight reasoning models, how chain-of-thought and reinforcement learning train them, test-time compute, open stacks, benchmarks and limits.

103topics
11Categories
5FAQs answered
01Proprietary reasoning models

OpenAI reasoning models

OpenAI's guide to its reasoning models. reasoning.effort accepts none, minimal, low, medium, high, xhigh and max depending on the model, and the GPT-5.6 and GPT-6 families add reasoning.mode (standard or pro). Raw reasoning tokens are not exposed; summaries are opt-in.

Best for: Hard coding, maths and planning tasks via API.

Visit OpenAI reasoning models →
02Proprietary reasoning models

Claude Fable 5.1

Anthropic's model for demanding reasoning and long-horizon agentic work. Adaptive thinking is always on, the default effort is high, with a 1M-token context window and 128K maximum output.

Best for: The hardest reasoning and long-running agent tasks.

Visit Claude Fable 5.1 →
03Proprietary reasoning models

Claude Opus 5.5

Anthropic's model for long-running agentic coding and knowledge work. Adaptive thinking is always on, the default effort is medium, with a 1M-token context window and 128K maximum output.

Best for: Agentic coding and knowledge work.

Visit Claude Opus 5.5 →
04Proprietary reasoning models

Claude Sonnet 5.5

Anthropic's balance of speed and intelligence. It uses adaptive thinking with a default effort of high, a 1M-token context window and 128K maximum output.

Best for: General reasoning with lower latency.

Visit Claude Sonnet 5.5 →
05Proprietary reasoning models

Claude Haiku 5.5

Anthropic's fastest model, for high-volume, latency-sensitive work such as classification, extraction and routing. It uses adaptive thinking with a default effort of medium and a 1M-token context window.

Best for: Fast, high-volume reasoning-light tasks.

Visit Claude Haiku 5.5 →
06Proprietary reasoning models

Gemini 4 Argon

Google's stable flagship reasoning model (gemini-4-argon) for software engineering, maths and science, and multi-step agents. Thinking levels are low, medium and high; the context window is 1,048,576 tokens, with long decode up to 1M output tokens.

Best for: Complex reasoning and agents on Gemini.

Visit Gemini 4 Argon →
07Proprietary reasoning models

Gemini 3.8 Flash

Google's most intelligent Flash model (gemini-3.8-flash), built for long-horizon software engineering and autonomous agents. Thinking levels are low, medium and high; minimal is not supported.

Best for: Faster, lower-cost Gemini reasoning.

Visit Gemini 3.8 Flash →
08Proprietary reasoning models

Gemini 3.8 Live Extended Thinking

Google's audio-to-audio Live API model (gemini-3.8-live-extended-thinking) for voice interactions that need extended background reasoning. Thinking levels are low, medium and high.

Best for: Voice agents that must reason before replying.

Visit Gemini 3.8 Live Extended Thinking →
09Reasoning modes & controls

OpenAI o1

OpenAI's first o-series reasoning model in the API. The deprecations page lists o1 and o1-pro for shutdown on 23 October 2026, with gpt-5.6-sol as the recommended replacement.

Best for: Understanding the o-series lineage and migrating off it.

Visit OpenAI o1 →
10Reasoning modes & controls

OpenAI o3-mini

Smaller o-series reasoning model. The deprecations page lists it for shutdown on 23 October 2026, with gpt-5.6-sol recommended.

Best for: Checking migration paths for low-cost reasoning.

Visit OpenAI o3-mini →
11Reasoning modes & controls

OpenAI o3

Reasoning model for maths, science, coding and visual reasoning, with a 200,000-token context window and 100,000 maximum output tokens. Marked deprecated; the deprecations page lists shutdown on 11 December 2026.

Best for: Reference specifications and migration planning.

Visit OpenAI o3 →
12Reasoning modes & controls

OpenAI o3-pro

Higher-compute variant of o3. Listed for shutdown on 11 December 2026; the deprecations page recommends gpt-5.6-sol with reasoning.mode set to pro.

Best for: Seeing how pro mode replaced a separate model.

Visit OpenAI o3-pro →
13Reasoning modes & controls

OpenAI o4-mini

Compact o-series model that supports detailed reasoning summaries. Listed for shutdown on 23 October 2026, with gpt-5.6-terra recommended.

Best for: Low-cost reasoning migration.

Visit OpenAI o4-mini →
14Reasoning modes & controls

OpenAI deprecations

Shutdown dates and recommended replacements for OpenAI models, including the o-series.

Best for: Checking whether a model is still available.

Visit OpenAI deprecations →
15Reasoning modes & controls

Claude extended thinking (legacy)

Manual extended thinking (thinking.type enabled with budget_tokens, minimum 1,024). Anthropic marks it deprecated on Claude Opus 4.6 and Sonnet 4.6 and not accepted on later models, which use adaptive thinking with an effort setting.

Best for: Migrating older code to adaptive thinking.

Visit Claude extended thinking (legacy) →
16Reasoning modes & controls

Claude thinking overview

How Claude's thinking works: adaptive thinking, the display setting (summarized or omitted) and thinking with tool use. The thinking text returned is a summary, never the raw chain of thought.

Best for: Understanding what Claude's reasoning output contains.

Visit Claude thinking overview →
17Reasoning modes & controls

Claude effort

output_config.effort sets how many tokens Claude spends: low, medium, high, xhigh for long-horizon work and max. Defaults differ by model, for example high on Claude Fable 5.1 and Sonnet 5.5 and medium on Opus 5.5 and Haiku 5.5.

Best for: Trading thoroughness against cost and speed.

Visit Claude effort →
18Reasoning modes & controls

Claude steering thinking and cost

Guidance on letting Claude decide when and how much to think, and on how thinking is priced.

Best for: Tuning thinking depth and cost.

Visit Claude steering thinking and cost →
19Reasoning modes & controls

Gemini thinking (API docs)

thinking_level support varies by model, for example low, medium and high on gemini-4-argon and gemini-3.8-flash. The docs also cover optional thought summaries and encrypted thought signatures that must be returned in stateless mode.

Best for: Setting Gemini thinking levels.

Visit Gemini thinking (API docs) →
20Reasoning modes & controls

Google Cloud: Gemini thinking

Google Cloud documentation on thinking for Gemini models, including thought signatures and a prompting guide.

Best for: Using Gemini thinking on Google Cloud.

Visit Google Cloud: Gemini thinking →
21Reasoning modes & controls

DeepSeek thinking mode (API guide)

Thinking is on by default at high effort and is set with thinking.type or reasoning_effort (low, high, max). The chain of thought comes back in reasoning_content and must be passed back when tools are used.

Best for: Calling DeepSeek's thinking models correctly.

Visit DeepSeek thinking mode (API guide) →
22Reasoning modes & controls

DeepSeek-R1-0528

Updated R1 checkpoint of 685B parameters under the MIT licence. The model card reports AIME 2025 accuracy rising from 70% to 87.5%, which are DeepSeek's own figures.

Best for: Running or studying the updated open R1 weights.

Visit DeepSeek-R1-0528 →
23Reasoning modes & controls

DeepSeek-R1 paper (arXiv)

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.

Best for: How R1 was trained.

Visit DeepSeek-R1 paper (arXiv) →
24Open-weight reasoning models

DeepSeek-R1

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 →
25Open-weight reasoning models

Qwen QwQ-32B

Alibaba's 32-billion-parameter open reasoning model trained with reinforcement learning.

Best for: Capable reasoning at a manageable size.

Visit Qwen QwQ-32B →
26Open-weight reasoning models

Kimi K2

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 →
27Open-weight reasoning models

Mistral Magistral Small

Mistral AI's Magistral Small reasoning model, with an emphasis on multilingual step-by-step reasoning.

Best for: European and multilingual reasoning.

Visit Mistral Magistral Small →
28Open-weight reasoning models

OpenAI gpt-oss

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 →
29Open-weight reasoning models

Phi-4-reasoning

Microsoft's small open reasoning model fine-tuned from Phi-4 for maths and science problems. GitHub lists v0.0.9 as the latest release, from January 2026.

Best for: Reasoning on limited hardware.

Visit Phi-4-reasoning →
30Open-weight reasoning models

Llama Nemotron Ultra

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 →
31Open-weight reasoning models

Olmo 3 Think

Ai2's fully open Olmo 3 Think reasoning model, released with data and training details.

Best for: Studying reasoning with open data and recipes.

Visit Olmo 3 Think →
32Open-weight reasoning models

DeepSeek-V4.1-Flash

DeepSeek's MIT-licensed V4.1-Flash, first published on 10 September 2026. Its model card describes a continuously controllable reasoning effort from 1 to 100.

Best for: Open-weight reasoning with a fine-grained effort control.

Visit DeepSeek-V4.1-Flash →
33Open-weight reasoning models

Kimi-K3

Moonshot AI's Kimi-K3, first published on 13 June 2026, is described on its model card as an open-weight, natively multimodal agentic model of about 2.8 trillion parameters. Its reported results use a reasoning-effort setting.

Best for: Very large open-weight reasoning and agent work.

Visit Kimi-K3 →
34Open-weight reasoning models

GLM-5.3

Z.ai's GLM-5.3, published on 25 August 2026. Its model card says the thinking budget is set with a reasoning_effort parameter with three levels: low, high and max.

Best for: Open-weight models with a three-level thinking budget.

Visit GLM-5.3 →
35Open-weight reasoning models

Qwen3.8-27B

Alibaba's Apache-2.0 Qwen3.8-27B, published in August 2026. Its model card says thinking mode is on by default, can be disabled per request, and reasoning depth can be tuned with reasoning_effort.

Best for: A compact open reasoning model with adjustable depth.

Visit Qwen3.8-27B →
41More reasoning models & provider docs

Kimi K2

Open-weight mixture-of-experts model from Moonshot AI.

Visit Kimi K2 →
42More reasoning models & provider docs

Olmo

Fully open models from Ai2 with open data and training code.

Visit Olmo →
59Training: verifiers, rewards & RL

Math-Shepherd

Process reward models without human step labels.

Visit Math-Shepherd →
61Training: verifiers, rewards & RL

DeepSeekMath

Introduced GRPO, a critic-free RL algorithm widely reused.

Visit DeepSeekMath →
67Training: verifiers, rewards & RL

Quiet-STaR

Models learn to think before speaking at every token.

Visit Quiet-STaR →
75Test-time compute & efficient reasoning

vLLM reasoning outputs

Serving reasoning models and parsing their thinking content.

Visit vLLM reasoning outputs →
77Open training stacks, libraries & datasets

Open-R1

Hugging Face's open reproduction of the DeepSeek-R1 training pipeline.

Best for: Learning how reasoning models are trained.

Visit Open-R1 →
78Open training stacks, libraries & datasets

s1: Simple test-time scaling

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 →
79Open training stacks, libraries & datasets

TRL

Hugging Face library for post-training with RL.

Visit TRL →
80Open training stacks, libraries & datasets

verl

Flexible RL training library for LLMs.

Visit verl →
81Open training stacks, libraries & datasets

OpenRLHF

Scalable RLHF and reasoning RL framework.

Visit OpenRLHF →
82Open training stacks, libraries & datasets

open-instruct (Ai2)

Ai2's post-training code, including RL with verifiable rewards.

Visit open-instruct (Ai2) →
83Open training stacks, libraries & datasets

OpenThoughts

Open reasoning datasets and models.

Visit OpenThoughts →
85Open training stacks, libraries & datasets

NuminaMath 1.5

Competition-level maths problems and solutions.

Visit NuminaMath 1.5 →
86Open training stacks, libraries & datasets

PRM800K

Step-level correctness labels for maths solutions.

Visit PRM800K →
87Benchmarks & evaluation

ARC Prize

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 →
88Benchmarks & evaluation

Humanity's Last Exam

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 →
89Benchmarks & evaluation

Epoch AI Benchmarks

Epoch AI's hub that tracks model results across benchmarks over time.

Best for: Comparing models across several benchmarks.

Visit Epoch AI Benchmarks →
90Benchmarks & evaluation

SWE-bench

Benchmark and leaderboard that tests models on resolving real GitHub issues in software repositories.

Best for: Reasoning applied to code.

Visit SWE-bench →
93Benchmarks & evaluation

GPQA

Graduate-level science questions.

Visit GPQA →
94Benchmarks & evaluation

MMLU-Pro

Harder multi-task knowledge and reasoning benchmark.

Visit MMLU-Pro →
101Limits, faithfulness & safety

The Illusion of Thinking

Studies how reasoning models behave as puzzle complexity rises; its conclusions have been debated.

Visit The Illusion of Thinking →

Reasoning Modes and Controls Compared

What each vendor's documentation says about turning reasoning up or down and what you can read back. These settings change quickly, so check each vendor's page for the model you use. Details here are as documented on 30 October 2026.

VendorHow you control reasoningWhat you can seeNotes
OpenAIreasoning.effort: none, minimal, low, medium, high, xhigh or max, depending on the model. reasoning.mode: standard or pro on the GPT-5.6 and GPT-6 families.Raw reasoning tokens are not exposed. Summaries are opt-in (auto, detailed or concise). Reasoning tokens are billed as output tokens and counted in output_tokens_details.reasoning_tokens.The o-series (o1, o3-mini, o3, o3-pro, o4-mini) is listed for shutdown between 23 October and 11 December 2026.
ClaudeAdaptive thinking (thinking.type adaptive), used by the current Claude 5.5 models, with output_config.effort: low, medium, high, xhigh or max. Manual extended thinking (thinking.type enabled with budget_tokens, minimum 1,024) is a legacy feature that Claude 4.7 and later models reject.The thinking text returned is a summary of the reasoning, never the raw chain of thought. display can be summarized or omitted; thinking is billed the same either way.Defaults differ by model; some models accept thinking.type disabled only at lower effort levels.
Geminithinking_level: minimal, low, medium or high, with each model supporting only some values (gemini-4-argon and gemini-3.8-flash accept low, medium and high). Dynamic thinking is the default.Thought summaries with thinking_summaries set to auto. Encrypted thought signatures must be resent unchanged in stateless mode.The docs do not describe a general off switch; Gemini 2.5 Flash-Lite defaults to thinking off.
DeepSeekthinking.type enabled or disabled, and reasoning_effort: low, high or max (default high). Thinking is on by default.The chain of thought is returned in reasoning_content and must be passed back when tools are used.temperature and the presence and frequency penalties have no effect in thinking mode. R1 itself dates from 20 January 2025 under the MIT licence.

Choosing the right approach

Match your requirement to a starting point, then compare vendors.

Hard maths, code or planningA proprietary reasoning model with a high effort or thinking budget
Control over data and costAn open-weight reasoning model such as DeepSeek-R1, gpt-oss or Qwen QwQ
Reasoning on small hardwareA distilled or compact model such as Phi-4-reasoning
Quick conversational repliesA standard LLM, since reasoning models add latency and cost
Know which reasoning control to setThe comparison above, then each vendor's thinking or reasoning guide
Learn how they are trainedOpen-R1 and the s1 research code
Judge claimsARC Prize, Humanity's Last Exam and your own test cases

Related pages on PeterIndia.net

Neighbouring directories and hubs.

Frequently asked questions

Quick answers.

What are reasoning LLMs and large reasoning models (LRMs)?

A reasoning LLM, also called a large reasoning model, generates an explicit chain of intermediate steps before its final answer and spends more compute at inference time on harder problems. It is typically trained with reinforcement learning on tasks whose answers can be checked, such as maths and code.

How are reasoning models trained?

Common ingredients are supervised fine-tuning on reasoning traces, reward signals from outcome verifiers or process reward models, and reinforcement learning with verifiable rewards using algorithms such as GRPO. DeepSeek-R1, Kimi k1.5, Tulu 3 and Qwen3 describe open recipes.

How is a reasoning model different from a standard LLM?

A standard LLM answers directly from what it learned. A reasoning model works through steps, checks its work and can use a larger thinking budget, at the cost of more latency and tokens. It helps most on maths, code, logic and planning, and can be slower and costlier without being better on simple lookups, writing or chat.

What is test-time compute scaling?

It means improving answers by using more computation at inference, for example longer chains of thought, sampling many answers and voting, or search guided by a verifier. Research such as Snell et al. and s1 shows it can rival scaling model size on some tasks.

Is the displayed reasoning a faithful explanation?

Not necessarily. Visible reasoning text may not reflect how the model reached its answer, and reasoning models can overthink simple questions. Verify conclusions independently, and test on your own tasks because benchmark scores date quickly.