DeepSeek 4x Faster: How Together AI's ATLAS System Works for LLMs
Together AI has introduced ATLAS, an adaptive speculator built on machine learning that speeds up LLM inference 4x with no manual tuning. The system learns automatically and adapts to your workload as you use it. On DeepSeek-V3.1 it reaches 500 tokens per second. That is 2.65 times faster than standard decoding, and it outperforms Groq's specialized hardware.
AI-processed from Together AI Blog; edited by Hamidun News
Together AI's ATLAS system can speed up inference for LLMs such as DeepSeek-V3.1 and Kimi-K2 by 4x, to 500 and 460 tokens per second respectively, with no manual tuning.
Which models does ATLAS work with?
ATLAS speeds up DeepSeek-V3.1 to 500 tokens per second and Kimi-K2 to 460 tokens per second, improving automatically as it is used.
What is speculative decoding?
Speculative decoding is one of the most powerful ways to speed up text generation in language models. In the standard approach, the model generates one token at a time; speculative decoding lets it process several tokens at once.
How much does ATLAS speed up DeepSeek?
ATLAS speeds up DeepSeek-V3.1 to 500 tokens per second, a 4x speedup. On Kimi-K2, the system reaches 460 tokens per second.
What results does ATLAS deliver on DeepSeek?
On DeepSeek-V3.1, ATLAS reaches 500 tokens per second (roughly a 4x speedup); on Kimi-K2, it reaches 460 tokens per second.
- Nov 1, 202616 models tested on 7 search tools: Exa averaged 10.5 out of 14, SearXNG about 3
- Nov 1, 2026MoneyGate safety layer blocked all 42 fraudulent requests to local AI models
- Oct 31, 2026DeepSeek-V4.1-Flash cuts HBM KV cache to 890 bytes per token, down from 3514
- Oct 30, 2026Mistral opens Large 4 weights: strong at code, but Chinese models score 32 to its 26
Key Questions Answered
Which models does ATLAS work with?
ATLAS speeds up DeepSeek-V3.1 to 500 tokens per second and Kimi-K2 to 460 tokens per second, improving automatically as it is used.
What is speculative decoding?
Speculative decoding is one of the most powerful ways to speed up text generation in language models. In the standard approach, the model generates one token at a time; speculative decoding lets it process several tokens at once.
How much does ATLAS speed up DeepSeek?
ATLAS speeds up DeepSeek-V3.1 to 500 tokens per second, a 4x speedup. On Kimi-K2, the system reaches 460 tokens per second.
What results does ATLAS deliver on DeepSeek?
On DeepSeek-V3.1, ATLAS reaches 500 tokens per second (roughly a 4x speedup); on Kimi-K2, it reaches 460 tokens per second.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.