RT @ArtificialAnlys: Cohere has launched Command A, a 111B parameter dense model that represents a huge leap from their previous Command R/….

Forgot to mention it the plot. The unit of time is in seconds.

Outlines is a fantastic open source project. It uses a novel idea to prebuild an FSM and pay the cost upfront and not during sampling. We used Outlines as a starting point and heavily replaced, reimplemented and optimized multiple components to arrive at a much faster solution.

The above benchmark was done a JSON schema with random keys and values as string.

Time to build JSON schema is proportinal to the tokenizer size. 256k is the tokenizer used for Cohere and Outlines. gpt-4o-mini was used for OpenAI so probably tokenizer size is 200k if not more as per tiktoken for gpt-4o.

Building artefacts for JSON Schema is expensive to do on scale in production for LLM inference. Probably why OpenAI had "JSON mode" for a year but not "JSON schema mode" until @cohere released it first :).* @cohere is >10x faster than OpenAI.* @cohere is >45x faster than Outlines

RT @sarahookr: Congrats to Sudip's team who released this a few weeks ago :). structured outputs are now in vogue -- and for good reason. A….

RT @DeanCarignan: Exciting to see more models supporting structured outputs. A big win for all developers who need to integrate LLMs into s….

RT @cohere: Response_format=json_object is here!. Command R models now support Structured Outputs for JSON. This forces the model to genera….

The thing I have been working on hard in the last few months at Cohere is finally out!!.

RT @Teknium1: Pretty incredible, grats @cohere.

RT @virattt: Cmd R+ beats Sonnet at financial RAG. I initially assumed these models were equivalent due to pricing. However, command r+ wa….

RT @cohere: Today, we’re introducing Command R+: a state-of-the-art RAG-optimized LLM designed to tackle enterprise-grade workloads and spe….

RT @aidangomez: ⌘-R. Introducing Command-R, a model focused on scalability, RAG, and Tool Use. We've also released the weights for research….

RT @KaustubhPriye: We are doing research on healthcare and Looking to connect with folks who have spent more than 50k out of their own pock….

I have a p-value joke, but I am not confident about it.

I have a joke about reading papers, but you want the Github repo first.

I have a deep learning joke, but it was already proposed by someone in the 1980s.

I have a Graph joke, but you won't be able to connect it.

I have an RNN joke, but I have an RNN joke.