Twitxy Saved

Together AI

This account is verified because it's an official organization on X. Learn more

@togethercompute

Accelerate inference, model shaping, and pre-training on a research-optimized platform.
San Francisco, CA together.ai Joined November 2022
279 Following 66K Followers 3.3K Posts

Latest posts 5

Pinned post

Together AI @togethercompute
Two weeks ago we announced Together Compiler: agents that tune inference to your model, chip, and traffic, for up to 2.6x more throughput.

Today, we’re sharing the journey behind it: the experiments that challenged our assumptions of megakernels, the surprising strengths of Torch Compile, and the lessons that shaped our development of a next-generation system for inference optimization.

Shout out to @dylan__lim @reyna_abhyankar @simran_s_arora and the team!

Read more: together.ai/blog/road-to-t…
13 replies 25 reposts 164 likes 60K views View on X
Together AI @togethercompute
Open models, put to the test on real research tasks. Thanks to @askalphaxiv for building the benchmark and running them on Together AI!
alphaXiv Verified account @askalphaxiv 8h
We created the world’s first open source benchmark for end-to-end autoresearch. Today, its results are live.

Which models are best for RSI? Can they understand the literature? Can they conduct complex scientific experiments, while maintaining good taste?

We spent about $7.4k in tokens across 8 models, 6 providers, and 3 evaluation tracks to find out.

The OpenResearch Intelligence Index evaluates models on scientific knowledge (20 hand-picked Terminal-Bench-Science tasks), post-training ability (6 novel post-training tasks), and literature relevancy/taste (a subset of PaperFindingBench), with each model running in its native harness inside OpenResearch.

So far, the index includes Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon, Fable 5.1, DeepSeek V4.1 Flash, Kimi K3, and Qwen3.8-Flash.

Opus 5.5 scores the highest at 61.8%, costing $16.46 per task. GPT-6.1 Sol scores slightly lower (54.9%) at a fraction of the cost (only $3.83 per task). When paired with OpenResearch’s literature search tools, DeepSeek V4.1 Flash performs nearly as well as GPT-6 Astra on the literature relevancy track (47.7% vs 48.7%, respectively).

Check out the blog post to learn more: alphaxiv.org/blog/introduci…. More models and source code coming soon.
View quoted post on X
1 replies 4 reposts 9 likes 1.4K views View on X
Together AI @togethercompute
together storage accelerator is now live, free with every together gpu cluster

it caches data from any s3-compatible object storage + hugging face on cluster nvme. one origin fetch per cluster, then warm reads up to 3.4 gb/s, ~20x faster than a cross-cloud pull

no code changes
0:46