Skynet Report

Yi-Coder, a new open-source code LLM series, has been released by 01.ai.

Available in 1.5B and 9B parameter versions, Yi-Coder offers base and chat models for efficient inference and flexible training.

The 9B version, built on Yi-9B, incorporates an additional 2.4T high-quality tokens from GitHub and filtered CommonCrawl data. Key features include continued pretraining on 52 major programming languages, a 128K token context window, and impressive performance metrics.

Yi-Coder-9B-Chat achieved a 23.4% pass rate on LiveCodeBench, outperforming larger models.

It also excelled in code editing, completion, and mathematical reasoning tasks, demonstrating capabilities comparable to or surpassing models with significantly more parameters.

Mistral AI has released Codestral, a 22B parameter code model designed for code generation.

It supports over 80 programming languages, including Python, Java, C, C++, JavaScript, and Bash.

Codestral features a 32k token context window, enabling it to outperform competitors on the RepoBench code completion evaluation.

It achieves notable results on benchmarks like HumanEval, MBPP, CruxEval, and Spider, demonstrating proficiency in Python and SQL. Codestral also excels in completing partial code segments across various languages.

It is available under the Mistral AI Non-Production License for research and testing, with commercial licenses available upon request.

IBM Research has released the Granite family of open source decoder-only code models ranging from 3 to 34 billion parameters.

The models are available in base and instruction-tuned variants optimised for enterprise software development workflows like code generation, fixing, explanation, and modernisation tasks.

Benchmarks show the Granite models match state-of-the-art performance among open source code LLMs across multiple coding tasks and programming languages.

The models leverage training data from sources like GitHub, CodeNet, and synthetic code-instruction pairs.

IBM is open sourcing the Granite code models under Apache 2.0 to enable open innovation and provide high-performing, cost-efficient foundation models for enterprises to build generative AI tools for developers.

Code Llama 70B is the largest and most advanced model in the Code Llama family.

This state-of-the-art language model is available in three versions: Code Llama – 70B, the foundational code model; Code Llama – 70B – Python, specialized for Python; and Code Llama – 70B – Instruct, fine-tuned for understanding natural language instructions.

Built on top of Llama 2, Code Llama is capable of generating code and natural language about code from both code and natural language prompts.

In internal benchmark testing, Code Llama outperformed other publicly available language models in code tasks and is free for both research and commercial use.

Magicoder is an open-source series of large language models (LLMs) for code developed to close the performance gap between synthetic instruction data-trained LLMs and real-world performance.

To do this, Magicoder was trained on 75k synthetic instruction data created by a method called OSS-Instruct, which uses open-source code snippets to generate instruction data.

This approach is designed to mitigate the inherent bias found in LLMs trained on synthetic data alone by enlightening the model with references, generating more diverse, realistic data and greater controllability.

Magicoder and its enhanced version, MagicoderS, outperform other code models on a variety of benchmarks, including text-to-code generation in Python, multilingual coding and data-science program completion.

Meta has released Code Llama, a large language model specialising in coding that it claims outperforms other publicly available options on code-related tasks.

Based on Meta’s LLM Llama 2 and built on a foundation of 500 billion code-related training tokens, Code Llama can generate code from prompts, as well as offering code completion and debugging support for a range of popular programming languages.

Three versions are being released, with 7, 13 and 34 billion parameters, with the 7B and 13B models also including “fill-in-the-middle” capabilities to support code completion.

While Meta recommends the use of Code Llama to assist software engineers, the company also warns developers not to use it for general natural language tasks, noting it is “not designed to follow natural language instructions”.