Models
Whether you're calling an API or running locally, the intelligence behind your app lives in a model. You want one that’s smart, fast, adaptable — and that can keep up as your use cases grow.
-
StarCoderOpen-access code model family from the BigCode project (Hugging Face and ServiceNow Research). StarCoder2 ships in 3B, 7B, and 15B sizes, trained on The Stack v2 across 600+ languages, with the 15B variant passing 70% on HumanEval.
-
Code LlamaMeta's open-source code model, built on Llama 2 for code generation and infilling. The repository was archived in July 2025 and is no longer maintained, superseded by newer Llama model generations.
-
DeepSeek
Chinese AI lab's open-weight model family: V3 for general tasks, R1 for reasoning, and a dedicated Coder line, all built on a mixture-of-experts architecture with a 128K context window. Code and weights are released under the MIT license and free to self-host or use commercially.
-
Phi-3 Code
Microsoft's small language model family, with Phi-3-mini (3.8B parameters) delivering GPT-3.5-level capability on a laptop or phone. Superseded by Phi-4 and the 2026 Phi-4-reasoning-vision-15B release, both open source under MIT.
-
ClaudeAnthropic's AI assistant, built for problem solving, data analysis, writing code, and working through complex tasks. The 2026 model lineup spans Claude Fable 5.1, Opus 5, Sonnet 5, and Haiku 4.5, with Claude Code for agentic coding included starting on the $20/month Pro plan.
-
Salesforce CodeGenOpen-source family of code-generation models from Salesforce Research, trained up to 16B parameters and competitive with early OpenAI Codex. The latest CodeGen2.5 matches larger 16B models at just 7B parameters.
-
Codestral
Mistral's 22B-parameter code model with a 256k context window and fill-in-the-middle completion, scoring 86.6% on HumanEval. Relicensed to Apache 2.0 in April 2026, making it free for commercial use.
