DeepSeek-R1 is a 671B parameter Mixture-of-Experts (MoE) model with 37B activated parameters per token, trained via large-scale reinforcement learning with a focus on reasoning capabilities. It incorporates two RL stages for discovering improved reasoning patterns and aligning with human preferences, along with two SFT stages for seeding reasoning and non-reasoning capabilities. The model achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks.
Over the past five months since the release of Qwen2-VL, developers have built new models based on it, contributing valuable feedback. Now, Qwen2.5-VL introduces enhanced capabilities, including precise analysis of images, text, and charts, as well as object localization with structured JSON outputs. It understands long videos, identifies key events, and functions as an agent, interacting with tools on computers and phones. The model's architecture features dynamic video processing and an optimized ViT encoder for improved speed and accuracy.
| DeepSeek-R1 | Qwen2.5-VL-32B | |
|---|---|---|
Web Site
| ||
Provider
| ||
Chat
| ||
Release Date
| ||
Modalities
| text | text images video |
API Providers
| DeepSeek, HuggingFace | - |
Knowledge Cut-off Date
| Unknown | Unknown |
Open Source
| Yes | Yes (Source) |
Pricing Input
| $0.55 per million tokens | $0 |
Pricing Output
| $2.19 per million tokens | $0 |
MMLU
| 90.8% Pass@1 Source | 78.4% Source |
MMLU-Pro
| 84% EM Source | 49.5% |
MMMU
| - | 70% |
HellaSwag
| - | Not available |
HumanEval
| - | Not available |
MATH
| - | 82.2% |
GPQA
| 71.5% Pass@1 Source | 46.0% Diamond |
IFEval
| 83.3% Prompt Strict Source | Not available |
SimpleQA
| - | - |
AIME 2024 | - | - |
AIME 2025 | - | - |
Aider Polyglot
| - | - |
LiveCodeBench v5
| - | - |
Global MMLU (Lite)
| - | - |
MathVista
| - | - |
Mobile Application | - | |
MathArena | ||
| Avg. Score | 82% | - |
AIME 2025 A test based on problems from the American Invitational Mathematics Examination, designed to assess the mathematical skills of models. | 89% | - |
HMMT February 2025 A test based on problems from the Harvard-MIT Mathematics Tournament, February 2025, designed to assess the mathematical skills of models. | 77% | - |
BRUMO 2025 | 92% | - |
SMT 2025 A test based on problems from the Stanford Math Tournament, 2025, designed to assess the mathematical skills of models. | 83% | - |
CMIMC 2025 A test based on problems from the Canadian Mathematical Olympiad, 2025, designed to assess the mathematical skills of models. | 69% | - |
Compare AI. Test. Benchmarks. Mobile Apps Chatbots, Sketch
Copyright © 2026 All Right Reserved.
Comments (1)
Jacquie
27 January 2026Digiturk paketler icinden size en uygun uyeligi kolayca secebilirsiniz. https://digiturkpaketler.com/kampanyalar