Code Models
AI-powered coding assistants for development
Code generation models for autocomplete, review, and refactors
Code-generation models are large language models trained or fine-tuned specifically on source code. They power IDE autocomplete, PR review, automated refactoring, test generation, and cross-language translation. Reach for a code model — over a general text model — when you want stronger correctness on programming tasks and structured outputs (diffs, JSON) that play well with developer tooling.
21 models available
Codestral
Mistral's code-specialized model. Optimized for code generation, completion, and understanding across 80+ languages.
Code Llama 13B Instruct
Meta's 13B Code Llama tuned for instruction following. A faster mid-size option for code generation and completion, supporting infilling for inserting code at a cursor position. Served on Replicate per call.
Code Llama 34B Instruct
Meta's 34B Code Llama tuned for instruction following. A balance of size and quality for code generation, completion, and explanation, with strong coverage of Python, JavaScript, and other common languages. Runs on Replicate per call.
Code Llama 70B Instruct
Meta's largest Code Llama, a 70B Llama-2 derivative specialized for programming and tuned to follow instructions in chat form. Handles code generation, completion, and explanation across common languages. Served on Replicate as a per-call endpoint.
Code Llama 7B Instruct
Meta's smallest Code Llama at 7B parameters, tuned for instruction following. The cheapest and fastest member of the family for quick code generation, completion, and infilling. Served on Replicate per call.
CodeGen 350M Mono
350M autoregressive code generation model from Salesforce, the smallest of the original CodeGen family. The mono variant was further trained on Python so it is well suited for short Python completions and program synthesis from a natural-language or code prompt.
DeepSeek Coder 1.3B Instruct
1.3B instruction-tuned code model from DeepSeek, trained on 2 trillion tokens of code and natural language across 87 languages with a 16k context window. One of the strongest tiny coders for its size, handling generation, completion and short coding instructions.
DeepSeek Coder 33B Instruct (GGUF)
Quantized GGUF build of DeepSeek's 33B code model, trained on roughly 2T tokens that are about 87 percent code. Designed for repository-level completion and project-aware generation thanks to a 16k context window. Runs on Replicate as a per-call endpoint.
DeepSeek Coder V2
DeepSeek's specialized coding model. Excellent at code generation, debugging, and explanation.
Granite Code 20B
IBM Granite 20B Code Instruct. Larger Granite code model balancing quality and inference cost for enterprise CI/CD code-review automation.
Granite Code 8B
IBM Granite 8B Code Instruct. Trained on permissively-licensed code, strong on multi-language code completion and instruction-following.
Grok Build 0.1
xAI's Grok coding-focused model. Tuned for code generation and software development tasks with a 256k token context window for working over large codebases.
Magicoder S CL 7B
UIUC Magicoder S CL 7B. CodeLlama-7B fine-tuned with OSS-Instruct synthetic data. Strong HumanEval Plus and MBPP Plus performance per parameter.
Phind CodeLlama 34B v2
Phind CodeLlama 34B v2. Highly tuned CodeLlama variant focused on retrieval-augmented developer assistant workflows.
Qwen2.5-Coder 32B Instruct
Alibaba's largest open Qwen2.5-Coder model. Trained on a code-heavy corpus, it matches or beats much larger general models on code generation and repair benchmarks like HumanEval and MBPP, and supports over 40 programming languages with fill-in-the-middle completion.
Qwen2.5-Coder 7B Instruct
The 7B instruct member of Alibaba's Qwen2.5-Coder family. A lighter, faster option for code completion, generation, and bug fixing across 40+ languages, with a 128k context and fill-in-the-middle support. Good price-to-quality balance for everyday coding tasks.
Replit Code v1 3B
Replit's 3B code-completion model, trained on a permissively licensed code subset of the Stack across 20 programming languages. Built for low-latency autocomplete rather than chat. Served on Replicate per call.
Replit Code v1.5 3B
3B code completion model from Replit trained on roughly 1 trillion tokens of permissively licensed code across 30 programming languages, with a 4k context window. Designed for autocomplete-style code generation and fill-in-the-middle.
Stable Code Instruct 3B
Instruction-tuned 3B code model from Stability AI, fine-tuned from stable-code-3b for chat-style coding tasks. Handles code generation, explanation and fix-up across multiple languages and was competitive with larger code models on benchmarks at release.
StarCoder2 15B
BigCode StarCoder2 15B code-generation flagship. Trained on 4T tokens of Stack v2 data with grouped-query attention and 16k context.
WizardCoder 33B
WizardLM WizardCoder 33B v1.1. Evol-Instruct fine-tune of DeepSeek-Coder-33B with strong code-generation benchmark performance.
Top code models picks
Hand-picked across four common criteria — resolved against the live catalog so the picks track price and performance changes.
Mistral's code-specialized model. Optimized for code generation, completion, and understanding across 80+ languages.
Learn more350M autoregressive code generation model from Salesforce, the smallest of the original CodeGen family. The mono variant was further trained on Python so it is well suited for short Python completions and program synthesis from a natural-language or code prompt.
Learn moreMistral's code-specialized model. Optimized for code generation, completion, and understanding across 80+ languages.
Learn moreMistral's code-specialized model. Optimized for code generation, completion, and understanding across 80+ languages.
Learn morePricing in code generation follows the same per-token model as general text. Flagship code models (GPT-5 Codex, Claude 4.6 Sonnet, Codestral) cost €1-€10 per million input tokens; budget tiers (Codestral Mamba, DeepSeek Coder, Qwen Coder) cost €0.05-€0.50 per million. A single IDE autocomplete request rarely runs more than a few thousand input tokens, so per-call cost is fractions of a cent. The bills grow when you ship agents that re-prompt themselves dozens of times per task.
The trade-off triangle is correctness, speed, and context. Flagships solve harder problems and follow project conventions more reliably but respond at 30-80 tokens/second, which feels slow inside a tight autocomplete loop. Fast budget models (Codestral Mamba, GPT-5 Mini) stream at 200+ tokens/second and feel native in the editor. For batch tasks (refactor a whole repo, generate tests for fifty files), flagship correctness wins. For tight autocomplete loops, fast tier wins.
Watch out for cross-file context: most autocomplete loops only send the current file. For real codebase-aware refactoring, you need a retrieval layer that pulls related files into the prompt. Tools like Cursor and Continue do this automatically; if you're rolling your own, embed the codebase first and retrieve the top 5-10 most relevant files per request.
Watch out for license contamination: a few open-weights code models were trained on permissively licensed code only; others swept GPL code with unclear redistribution terms. If you're shipping generated code in a closed-source product, prefer commercial models with explicit code-license guarantees.
Top picks above cover the most correct flagship, the cheapest workhorse, the longest-context model, and the fastest autocomplete option.
Popular use cases
Common patterns built with code models on Railwail.
Frequently asked questions
Start Building with AI
Access all models through a single API. Get free credits when you sign up — no credit card required.