Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official1
minimaxi.com2025-01-15officialabout the MiniMax 01 familyHN 1 pts · 0 comments
Analysis3
github.com2025-01-14authorityabout the MiniMax 01 familyHN 19 pts · 1 comments2,423 words
MiniMax-AI releases MiniMax-Text-01 and MiniMax-VL-01, large language and vision-language models using Linear Attention. The text model has 456 billion total parameters with 45.9 billion activated per token and supports a 1 million token training context length extending to 4 million tokens during inference. The vision model uses a 303-million-parameter Vision Transformer and recommends int8 quantization for deployment.written, reviewed, cited
- MiniMax-Text-01 positions a softmax attention layer after every 7 lightning attention layers.
- The model employs a Top-2 routing strategy to select experts within its Mixture-of-Experts architecture.
- Deployment requires converting specific MoE gate coefficients and heads to int8 while protecting the lm_head and embed_tokens.
arxiv.org2025-01-15authorityabout the MiniMax 01 familyHN 3 pts · 1 comments365 words
MiniMax-01 releases MiniMax-Text-01 and MiniMax-VL-01 using lightning attention and a MoE architecture with 32 experts and 456 billion total parameters, where 45.9 billion are activated per token. MiniMax-Text-01 trains on contexts up to 1 million tokens and extrapolates to 4 million tokens during inference, while MiniMax-VL-01 utilizes 512 billion vision-language tokens. Benchmarks show these models match GPT-4o and Claude-3.5-Sonnet performance with 20-32 times longer context windows.written, reviewed, cited
yacinemahdid.com2025-04-01about the MiniMax 01 familyHN 2 pts · 0 comments1,151 words
Minimax-01 is an MIT-licensed, 460 billion parameter model that integrates Lightning Attention with a Mixture of Experts to support context windows up to 4 million tokens. The architecture hybridizes attention mechanisms using an 8:1 ratio of linear to softmax blocks, enabling efficient scaling while matching closed-source model performance. This approach allows the system to extrapolate inference capabilities significantly beyond its training window of 1 million tokens.written, reviewed, cited
- Linear attention models struggle to match softmax performance in tasks requiring complex reasoning or long-range dependencies.
- Lightning Attention minimizes cumsum operations by adopting a tiling strategy similar to FlashAttention to overcome training inefficiency.
- The model requires a specific recipe detailed in sections 4 and 5 of the paper to achieve its reported scaling and performance.
