Reasoning via Large-Scale RL

Reinforcement learning on outcomes, often verifiable math or code, to produce long reasoning traces.

Tier Grades

Tier A: Frontier AI labs

feasible

high confidence

OpenAI, xAI, DeepSeek, MiniMax and Mistral are observed, with Grok 4 RL at 200,000-GPU pretraining scale.

Tier B: Indonesian AI labs

stretch

medium confidence

The cheapest anchors are $294,000 to $534,700 for the RL stage alone on top of a strong base (MiniMax used 512 H800 for 3 weeks), and the notes verify no Indonesian national compute programme, BRIN facility or ITB GPU inventory (T11 Part E item 2).

Tier C: Institution units such as ITB

not feasible

high confidence

512 H800 for 3 weeks far exceeds the typical academic 1 to 8 GPUs, and 85% of surveyed academics have no cloud budget.

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

not feasible

medium confidence

The notes give no anchor for proof-of-concept firms at hundreds of GPUs for weeks.

Where this comes from

Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.

Anchors

ComputeGrok 4 RL at pretraining scale on a 200,000 GPU cluster; MiniMax-M1 RL $534,700 on 512 H800 for about 3 weeks; DeepSeek-R1 RL stage $294,000 on top of an about $6M base model.
DataVerifiable rewards; R1 states reasoning can be incentivised through pure RL without human-labelled trajectories.
Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): not sourced in the notes.
Time horizonWeeks per RL run.
Shelf-life signalso1 performance improves smoothly with train-time and test-time compute
Pipeline roleTier A produces reasoning models and open reasoning checkpoints (R1, Magistral Small) that Tier B and Tier C distil or fine-tune and Tier D deploys.

Where this comes from

Team size is a proxy (author counts of exemplar papers), not a headcount.

Who and What

Exemplars

DeepSeek-R1 SpaceXAI (xAI)

Grok 4

Magistral

Observed actors

OpenAI

SpaceXAI (xAI) DeepSeek

DeepSeek

MiniMax

Mistral AI

Linked models

+50

Jamba Reasoning 3Breasoningproduced by

Ring-flash-2.0reasoningproduced by

DeepSeek-V3.1LLMproduced by

NVIDIA Nemotron Nano 9B v2reasoningproduced by

GPT-5 nanoreasoningproduced by

Claude Opus 4.1LLMproduced by

Command A Reasoning (08-2025)reasoningproduced by

EXAONE 4.0 (32B)reasoningproduced by

Phi-4-mini-flash-reasoningreasoningproduced by

MiniMax-M1reasoningproduced by

Magistral Smallreasoningproduced by

o3-proreasoningproduced by

Seed1.6-Thinkingreasoningproduced by

Claude Opus 4LLMproduced by

Absolute Zeroreasoningproduced by

Qwen3-235B-A22BLLMproduced by

Qwen3-30B-A3BLLMproduced by

Qwen3-32BLLMproduced by

o3reasoningproduced by

o4-minireasoningproduced by

Kimi-VL-A3B-Thinkingmultimodal LLMproduced by

Llama 3.1 Nemotron Ultra 253B v1reasoningproduced by

o1-proreasoningproduced by

Llama 3.3 Nemotron Super 49B v1reasoningproduced by

Llama-3.1-Nemotron-Nano-8B-v1reasoningproduced by

Reka Flash 3reasoningproduced by

Sonar Reasoning Proreasoningproduced by

QwQ-32Breasoningproduced by

TxAgent T1 Llama 3.1 8Breasoningproduced by

ERNIE X1reasoningproduced by

o3-minireasoningproduced by

HuatuoGPT-o1-72Breasoningproduced by

QVQ-72B-Previewreasoningproduced by

DeepSeek-V2.5 (Dec 2024, DeepSeek-V2.5-1210)LLMproduced by

QwQ-32B-Previewreasoningproduced by

o1-minireasoningproduced by

Command R (Mar '24)LLMproduced by

Olmo 3.1 32B Thinkreasoningproduced by

Nova 2 Litereasoningproduced by

ERNIE 5.0 Thinking Previewreasoningproduced by

LFM2.5-Thinking 1.2Bedge LLMproduced by

Phi-4-mini-reasoningreasoningproduced by

Phi-4-reasoningreasoningproduced by

Magistral Medium 1.2reasoningproduced by

Magistral Small 1.2reasoningproduced by

Kimi K2 Thinkingreasoningproduced by

HyperCLOVA X SEED Think 32Breasoningproduced by

GPT-6LLMproduced by

o-seriesLLMproduced by

MiMo-V2.6-Pro-RLLLMproduced by

Papers

Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsJason Wei et al., arXiv preprint, 2022 founding
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement LearningDeepSeek-AI et al., Nature volume 645, pages 633-638 (2025), 2025 state of the art
DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language ModelsZhihong Shao et al., arXiv preprint, 2024 state of the art
Let's Verify Step by StepHunter Lightman et al., arXiv preprint, 2023 state of the art
Towards Large Reasoning Models: A Survey of Reinforced Reasoning with Large Language ModelsFengli Xu et al., arXiv preprint, 2025 survey

Where this comes from

The papers to read first: where the direction began, its current best, and a survey when one exists.

Sources

Grok 4 x.ai https://x.ai/news/grok-4
DeepSeek on the Cover of "Nature": First Response to Controversies Led by LIANG Wenfeng eu.36kr.com https://eu.36kr.com/en/p/3471852231792260
MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention arxiv.org https://arxiv.org/abs/2506.13585
$100K or 100 Days:Trade-offs when Pre-Training with Academic Resources arxiv.org https://arxiv.org/html/2410.23261v1
Learning to reason with LLMs openai.com https://openai.com/index/learning-to-reason-with-llms/
Magistral | Mistral AI mistral.ai https://mistral.ai/news/magistral
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning arxiv.org https://arxiv.org/abs/2501.12948

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Reasoning via Large-Scale RL", https://aurion.semai.id/topics/reasoning-via-large-scale-rl/, records of October 2026.

Reasoning via Large-Scale RL

Tier Grades

Tier A: Frontier AI labs

Tier B: Indonesian AI labs

Tier C: Institution units such as ITB

Tier D: Organisations such as SEMAI (PoT, PoC, PoV)

Anchors

Who and What

Links

Models 74

Papers

Sources