1
Mamba-3
Meet Mamba-3: the SSM built for inference. Faster than Transformers at decode, stronger than Mamba-2, and open-source from day one.
Together AI Blog
Build what's next on the AI Native Cloud. Full-stack AI platform for inference, fine-tuning, and GPU clusters — powered by cutting-edge research.
- Popular posts
- 8
- Total points
- 1,769
- HN comments
- 434
- Rank
- #701 company
Visit together.ai → RSS feedOn Hacker News Oct 2023 – Mar 2026
AI & Machine Learning 8Systems & Performance 4Programming Languages 1Databases & Data 1
2
FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-Precision
FlashAttention-3 achieves up to 75% GPU utilization on H100s, making AI models up to 2x faster and enabling efficient processing of longer text inputs. It allows for faster training and inference of LLMs, supports lower precision operations for improved efficiency.
3
RedPajama v2 Open Dataset with 30T Tokens for Training LLMs
RedPajama-Data-v2 is a 30 trillion token open dataset from 84 CommonCrawl dumps across five languages, with over 40 precomputed quality annotations.
4
Paving the way to efficient architectures: StripedHyena-7B
StripedHyena-7B is an open-source model built beyond the Transformer architecture, matching Llama-2 and Mistral 7B while running faster on long sequences.
5
Consistency diffusion language models: Up to 14x faster, no quality loss
Standard diffusion language models can't use KV caching and need too many refinement steps to be practical. CDLM fixes both with a post-training recipe that enables exact block-wise KV caching and trajectory-consistent step reduction — delivering up to 14.5x latency improvements
6
AdapTive-LeArning Speculator System (ATLAS): Faster LLM inference
LLM inference that gets faster as you use it. Our runtime-learning accelerator adapts continuously to your workload, delivering 500 TPS on DeepSeek-V3.1, a 4x speedup over baseline performance without manual tuning.
7
Based: Simple linear attention language models
BASED is a linear attention architecture that decodes without a KV-cache, delivering a 24x throughput improvement over Transformers with FlashAttention-2.
8
Dragonfly: A large vision-language model with multi-resolution zoom
Dragonfly is a vision-language architecture with multi-resolution zoom for fine-grained image understanding, including a medical model beating Med-Gemini.