YOU GUYS — Cohere just dropped their first LLM dedicated entirely to code generation, called the NORTH-MINI model and it's genuinely impressive for its size! While my previous post said 30B parameters that was a complete brain fart because it is actually a 7B parameter model trained on 25 billion tokens of high-quality code across 63 different programming languages in what they call their 'CodeMaster' dataset. They even included FlashAttention v4 and Grouped-Query Attention to maximize throughput, which means this thing is built for speed as well as accuracy — I want you all to look at the benchmark numbers because that's where it gets wild! It scores 85% on HumanEval (that beating StarCoder1B/1.3B models nearly twice its size) and roughly a 60% pass rate on MBPP, which is incredibly strong for this class of model. They also ran the SWE-bench leaderboard and it held up against models double its parameter count in code generation tasks — that's not accidental; they spent months curating high-quality code instead of just scraping the whole internet like everyone else does!
The real magic though is what developers can actually do with this thing on a day-to-day basis. You don't have to just use it as an autocomplete engine in VS Code — you can feed it full function descriptions from comments, generate entire modules from spec docs, or use their fine-tuned encoder model for embedding-based code search that would beat almost anything else under 10B parameters. And since it supports MQA quantization and was built with FlashAttention v4, this is one of the first models I'd actually run locally on my workstation instead of constantly calling a paid API - you can download it directly as CohereLabs/northmini-code-1.0 from Hugging Face using transformers, or fine-tune your own version using their provided recipe which includes data cleaning and loss function weighting techniques they spent 9 months refining! Seriously — if you write code for a living this is one of the most practical new model releases I've seen in years because it solves real developer problems rather than just being another leaderboard number.
Source: https://huggingface.co/blog/CohereLabs/introducing-north-mini-code
The real magic though is what developers can actually do with this thing on a day-to-day basis. You don't have to just use it as an autocomplete engine in VS Code — you can feed it full function descriptions from comments, generate entire modules from spec docs, or use their fine-tuned encoder model for embedding-based code search that would beat almost anything else under 10B parameters. And since it supports MQA quantization and was built with FlashAttention v4, this is one of the first models I'd actually run locally on my workstation instead of constantly calling a paid API - you can download it directly as CohereLabs/northmini-code-1.0 from Hugging Face using transformers, or fine-tune your own version using their provided recipe which includes data cleaning and loss function weighting techniques they spent 9 months refining! Seriously — if you write code for a living this is one of the most practical new model releases I've seen in years because it solves real developer problems rather than just being another leaderboard number.
Source: https://huggingface.co/blog/CohereLabs/introducing-north-mini-code