The Phind team announced the release of a new AI model based on CodeLlama-34B. This model, which Phind now uses as standard, is capable of outperforming GPT-4 in terms of coding while generating answers to technical questions five times faster.
We are pleased to announce that Phind now defaults to our own model that matches and exceeds the encoding capabilities of GPT-4 while running five times faster. You can now receive high-quality answers to technical questions in 10 instead of 50 seconds.
The current 7th generation Phind model is based on our open source CodeLlama-34B optimizations, which were the first models to surpass GPT-4’s score on HumanEval and remain by far the best open source coding models in the world are margin.
- The Phind V7 model achieves 74.7% Pass@1 on HumanEval
This new model has been refined using more than 70 billion additional tokens for high-quality code and reasoning problems and has a HumanEval score of 74.7%. However, we found that HumanEval is a poor indicator of practical utility. After deploying previous iterations of the Phind model in our service, we collected detailed feedback and found that in most cases our model meets or exceeds the usefulness of GPT-4 on real-world questions. Many members of our Discord community have started using Phind exclusively with the Phind template, although they also have unlimited GPT-4 access.
One of the main advantages of the Phind model is that it is very fast. We were able to increase speed by 5x compared to GPT-4 by running our model on H100s with NVIDIA’s new TensorRT-LLM library, achieving 100 tokens per second in a single stream.
Another important advantage of the Phind model is context – it supports up to 16,000 tokens. We currently allow 12,000 tokens to be entered into the site and reserve the remaining 4,000 tokens for site results.
The Phind model still has some shortcomings and we will continue to improve it. One area where it still suffers is consistency – for some difficult questions where it gets the correct answer, the Phind model may take more generations to find the correct answer than GPT-4 .
Source : Phind
And you ?
What is your opinion on this topic?
Do you think the claim that Phind beats GPT-4 in terms of coding is credible?
What is your favorite AI tool for coding? GPT4? Co-pilot? Phind? Nobody? Other ?
See also
Meta is launching an AI tool called Code Llama that can generate computer code written by a programmer and debug code. But critics say the tools are unreliable
Meta makes its artificial intelligence and advanced language model Llama 2 available for commercial use through partnerships with Microsoft