Atlas

Models

← All models

K2 Think V2

LLM360 · 2026-01-27 · 72.6B parameters

K2 Think V2 is LLM360's 70B open-weight general reasoning model, built on K2-V2-Instruct and post-trained with two stages of reinforcement learning with verifiable rewards (RLVR) using GRPO. Its chat template defaults to high reasoning effort, while low and medium settings remain available but were not evaluated in the official release. The checkpoint supports up to 262,144 tokens through 2× YaRN scaling from a 131,072-token serving context, and official evaluations report strong results on AIME 2025, HMMT 2025, and GPQA-Diamond.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo58.8
AA-LCR52.7
AA-Omniscience Index-33.9
Artificial Analysis Agentic Index1.8
Artificial Analysis Coding Index21.0
Artificial Analysis Intelligence Index17.3
Artificial Analysis Omniscience Accuracy15.7
Artificial Analysis Omniscience Hallucination Rate58.9
Artificial Analysis Openness Index88.9
Capability136.7
CritPt0.0
GDPval-AA v2381
GDPval-AA v2 (normalized)0.0
GPQA Diamond71.3
Humanity's Last Exam (Text-Only)9.5
IFBench (Artificial Analysis)62.8
SciCode (Artificial Analysis)33.0
Terminal-Bench 2.115.0
Terminal-Bench Hard6.8
τ²-bench (Telecom)25.4
𝜏³-Banking5.4
Loading Atlas data…