Code Models

AI-powered coding assistants for development

Modèles de génération de code pour autocomplete, review et refactoring

Les modèles de génération de code sont des grands modèles de langage entraînés ou affinés spécifiquement sur du code source. Ils alimentent l'autocomplete IDE, la PR review, le refactoring automatisé, la génération de tests et la traduction cross-langage. On y a recours — plutôt qu'à un modèle texte généraliste — quand on veut une correction plus forte sur les tâches de programmation et des sorties structurées (diffs, JSON) qui se marient bien avec l'outillage dev.

21 models available

Codestral

CodeMistral AI
NewPopular

Mistral's code-specialized model. Optimized for code generation, completion, and understanding across 80+ languages.

Free1.5s
codingfastmultilanguage

Code Llama 13B Instruct

CodeMeta

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.

€1.00
metacode-llamacoding

Code Llama 34B Instruct

CodeMeta

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.

€2.00
metacode-llamacoding

Code Llama 70B Instruct

CodeMeta

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.

€3.00
metacode-llamacoding

Code Llama 7B Instruct

CodeMeta

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.

€1.00
metacode-llamacoding

CodeGen 350M Mono

Codehuggingface

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.

Free
codehuggingfacesalesforce

DeepSeek Coder 1.3B Instruct

Codehuggingface

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.

Free
codehuggingfacedeepseek

DeepSeek Coder 33B Instruct (GGUF)

CodeReplicate

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.

€2.00
deepseekcodinginstruct

DeepSeek Coder V2

CodeDeepSeek

DeepSeek's specialized coding model. Excellent at code generation, debugging, and explanation.

Free2.0s
codingaffordable

Granite Code 20B

CodeReplicate

IBM Granite 20B Code Instruct. Larger Granite code model balancing quality and inference cost for enterprise CI/CD code-review automation.

€0.006
replicatecode-generationibm

Granite Code 8B

CodeReplicate

IBM Granite 8B Code Instruct. Trained on permissively-licensed code, strong on multi-language code completion and instruction-following.

€0.004
replicatecode-generationibm

Grok Build 0.1

CodexAI

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.

Free
xaigrokcode

Magicoder S CL 7B

CodeCommunity

UIUC Magicoder S CL 7B. CodeLlama-7B fine-tuned with OSS-Instruct synthetic data. Strong HumanEval Plus and MBPP Plus performance per parameter.

€0.003
replicatecode-generationopen-weights

Phind CodeLlama 34B v2

CodeReplicate

Phind CodeLlama 34B v2. Highly tuned CodeLlama variant focused on retrieval-augmented developer assistant workflows.

€0.009
replicatecode-generationphind

Qwen2.5-Coder 32B Instruct

Codehuggingface

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.

Free
qwenalibabacoding

Qwen2.5-Coder 7B Instruct

Codehuggingface

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.

Free
qwenalibabacoding

Replit Code v1 3B

CodeReplicate

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.

€1.00
replitcodingcompletion

Replit Code v1.5 3B

Codehuggingface

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.

Free
codehuggingfacereplit

Stable Code Instruct 3B

Codehuggingface

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.

Free
codehuggingfacestability-ai

StarCoder2 15B

CodeCommunity

BigCode StarCoder2 15B code-generation flagship. Trained on 4T tokens of Stack v2 data with grouped-query attention and 16k context.

€0.005
replicatecode-generationbigcode

WizardCoder 33B

CodeCommunity

WizardLM WizardCoder 33B v1.1. Evol-Instruct fine-tune of DeepSeek-Coder-33B with strong code-generation benchmark performance.

€0.009
replicatecode-generationwizardlm

Top code models picks

Hand-picked across four common criteria — resolved against the live catalog so the picks track price and performance changes.

Meilleur global
Codestral

Mistral's code-specialized model. Optimized for code generation, completion, and understanding across 80+ languages.

Learn more
Le moins cher
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.

Learn more
Contexte le plus long
Codestral

Mistral's code-specialized model. Optimized for code generation, completion, and understanding across 80+ languages.

Learn more
Le plus rapide
Codestral

Mistral's code-specialized model. Optimized for code generation, completion, and understanding across 80+ languages.

Learn more

La tarification en génération de code suit le même modèle au token que le texte généraliste. Les modèles phares de code (GPT-5 Codex, Claude 4.6 Sonnet, Codestral) coûtent 1 à 10 € par million de tokens d'entrée ; les tiers économiques (Codestral Mamba, DeepSeek Coder, Qwen Coder) coûtent 0,05 à 0,50 € par million. Une seule requête d'autocomplete IDE dépasse rarement quelques milliers de tokens d'entrée, alors le coût par appel se compte en fractions de centime. Les factures gonflent quand vous livrez des agents qui se relancent des dizaines de fois par tâche.

Le triangle de compromis est correction, vitesse et contexte. Les phares résolvent des problèmes plus durs et suivent les conventions du projet plus fidèlement mais répondent à 30-80 tokens/seconde, ce qui paraît lent dans une boucle d'autocomplete serrée. Les modèles rapides économiques (Codestral Mamba, GPT-5 Mini) streament à 200+ tokens/seconde et paraissent natifs dans l'éditeur. Pour les tâches par lots (refactor de tout un repo, génération de tests pour cinquante fichiers), la correction phare gagne. Pour les boucles d'autocomplete serrées, le tiers rapide gagne.

Attention au contexte cross-fichier : la plupart des boucles d'autocomplete n'envoient que le fichier courant. Pour du refactoring conscient du codebase, vous avez besoin d'une couche de retrieval qui tire les fichiers connexes dans le prompt. Des outils comme Cursor et Continue le font automatiquement ; si vous le construisez vous-même, embeddez le codebase d'abord et récupérez les 5 à 10 fichiers les plus pertinents par requête.

Attention à la contamination de licence : quelques modèles open-weights ont été entraînés uniquement sur du code sous licence permissive ; d'autres ont aspiré du code GPL avec des termes de redistribution flous. Si vous livrez du code généré dans un produit closed-source, préférez les modèles commerciaux avec garanties explicites de licence de code.

Les top picks ci-dessus couvrent le phare le plus correct, le cheval de trait le moins cher, le modèle au plus long contexte et l'option autocomplete la plus rapide.

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.