Telegram Group & Telegram Channel
Forwarded from Machinelearning
🌟MiniMax-M1: открытя reasoning‑LLM с контекстом 1M

MiniMax-M1 — первая в мире open-weight гибридная reasoning‑LLM c 1M контекстом (8× DeepSeek R1) и гибридной архитектурой MoE + lightning attention.
• 456 млрд параметров (45,9 млрд активируются на токен), сверхэффективная генерация — 25% FLOPs DeepSeek R1 на 100K токенов
• Обучение через RL с новым алгоритмом CISPO, решающим реальные задачи от математики до кодинга
• На обучение было потрачено $534K, две версии — 40K/80K “thinking budget”
• Обходит DeepSeek R1 и Qwen3-235B на бенчмарках по математике и кодингу,
• Топ результат на задачах для software engineering и reasoning



Бенчмарки:
AIME 2024: 86.0 (M1-80K) vs 85.7 (Qwen3) vs 79.8 (DeepSeek R1)

SWE-bench Verified: 56.0 vs 34.4 (Qwen3)

OpenAI-MRCR (128k): 73.4 vs 27.7 (Qwen3)

TAU-bench (airline): 62.0 vs 34.7 (Qwen3)

LongBench-v2: 61.5 vs 50.1 (Qwen3)


➡️ Попробовать можно здесь

Hugging Face: https://huggingface.co/collections/MiniMaxAI/minimax-m1-68502ad9634ec0eeac8cf094
GitHub: https://github.com/MiniMax-AI/MiniMax-M1
Tech Report: https://github.com/MiniMax-AI/MiniMax-M1/blob/main/MiniMax_M1_tech_report.pdf


@ai_machinelearning_big_data

#llm #reasoningmodels #minimaxm1
Please open Telegram to view this post
VIEW IN TELEGRAM



group-telegram.com/machinelearning_interview/1863
Create:
Last Update:

🌟MiniMax-M1: открытя reasoning‑LLM с контекстом 1M

MiniMax-M1 — первая в мире open-weight гибридная reasoning‑LLM c 1M контекстом (8× DeepSeek R1) и гибридной архитектурой MoE + lightning attention.
• 456 млрд параметров (45,9 млрд активируются на токен), сверхэффективная генерация — 25% FLOPs DeepSeek R1 на 100K токенов
• Обучение через RL с новым алгоритмом CISPO, решающим реальные задачи от математики до кодинга
• На обучение было потрачено $534K, две версии — 40K/80K “thinking budget”
• Обходит DeepSeek R1 и Qwen3-235B на бенчмарках по математике и кодингу,
• Топ результат на задачах для software engineering и reasoning



Бенчмарки:
AIME 2024: 86.0 (M1-80K) vs 85.7 (Qwen3) vs 79.8 (DeepSeek R1)

SWE-bench Verified: 56.0 vs 34.4 (Qwen3)

OpenAI-MRCR (128k): 73.4 vs 27.7 (Qwen3)

TAU-bench (airline): 62.0 vs 34.7 (Qwen3)

LongBench-v2: 61.5 vs 50.1 (Qwen3)


➡️ Попробовать можно здесь

Hugging Face: https://huggingface.co/collections/MiniMaxAI/minimax-m1-68502ad9634ec0eeac8cf094
GitHub: https://github.com/MiniMax-AI/MiniMax-M1
Tech Report: https://github.com/MiniMax-AI/MiniMax-M1/blob/main/MiniMax_M1_tech_report.pdf


@ai_machinelearning_big_data

#llm #reasoningmodels #minimaxm1

BY Machine learning Interview





Share with your friend now:
group-telegram.com/machinelearning_interview/1863

View MORE
Open in Telegram


Telegram | DID YOU KNOW?

Date: |

In the United States, Telegram's lower public profile has helped it mostly avoid high level scrutiny from Congress, but it has not gone unnoticed. In addition, Telegram's architecture limits the ability to slow the spread of false information: the lack of a central public feed, and the fact that comments are easily disabled in channels, reduce the space for public pushback. Some people used the platform to organize ahead of the storming of the U.S. Capitol in January 2021, and last month Senator Mark Warner sent a letter to Durov urging him to curb Russian information operations on Telegram. However, the perpetrators of such frauds are now adopting new methods and technologies to defraud the investors. This ability to mix the public and the private, as well as the ability to use bots to engage with users has proved to be problematic. In early 2021, a database selling phone numbers pulled from Facebook was selling numbers for $20 per lookup. Similarly, security researchers found a network of deepfake bots on the platform that were generating images of people submitted by users to create non-consensual imagery, some of which involved children.
from tr


Telegram Machine learning Interview
FROM American