Best AI Coding Models for Linux - Page 2

DeepCoder is a fully open source code-reasoning and generation model released by Agentica Project in collaboration with Together AI. It is fine-tuned from DeepSeek-R1-Distilled-Qwen-14B using distributed reinforcement learning, achieving a 60.6% accuracy on LiveCodeBench (representing an 8% improvement over the base), a performance level that matches that of proprietary models such as o3-mini (2025-01-031 Low) and o1 while using only 14 billion parameters. It was trained over 2.5 weeks on 32 H100 GPUs with a curated dataset of roughly 24,000 coding problems drawn from verified sources (including TACO-Verified, PrimeIntellect SYNTHETIC-1, and LiveCodeBench submissions), each problem requiring a verifiable solution and at least five unit tests to ensure reliability for RL training. To handle long-range context, DeepCoder employs techniques such as iterative context lengthening and overlong filtering.

DeepSWE is a fully open source, state-of-the-art coding agent built on top of the Qwen3-32B foundation model and trained exclusively via reinforcement learning (RL), without supervised finetuning or distillation from proprietary models. It is developed using rLLM, Agentica’s open source RL framework for language agents. DeepSWE operates as an agent; it interacts with a simulated development environment (via the R2E-Gym environment) using a suite of tools (file editor, search, shell-execution, submit/finish), enabling it to navigate codebases, edit multiple files, compile/run tests, and iteratively produce patches or complete engineering tasks. DeepSWE exhibits emergent behaviors beyond simple code generation; when presented with bugs or feature requests, the agent reasons about edge cases, seeks existing tests in the repository, proposes patches, writes extra tests for regressions, and dynamically adjusts its “thinking” effort.

GLM-4.6V is a state-of-the-art open source multimodal vision-language model from the Z.ai (GLM-V) family designed for reasoning, perception, and action. It ships in two variants: a full-scale version (106B parameters) for cloud or high-performance clusters, and a lightweight “Flash” variant (9B) optimized for local deployment or low-latency use. GLM-4.6V supports a native context window of up to 128K tokens during training, enabling it to process very long documents or multimodal inputs. Crucially, it integrates native Function Calling, meaning the model can take images, screenshots, documents, or other visual media as input directly (without manual text conversion), reason about them, and trigger tool calls, bridging “visual perception” with “executable action.” This enables a wide spectrum of capabilities; interleaved image-and-text content generation (for example, combining document understanding with text summarization or generation of image-annotated responses).

GLM-4.1V is a vision-language model, providing a powerful, compact multimodal model designed for reasoning and perception across images, text, and documents. The 9-billion-parameter variant (GLM-4.1V-9B-Thinking) is built on the GLM-4-9B foundation and enhanced through a specialized training paradigm using Reinforcement Learning with Curriculum Sampling (RLCS). It supports a 64k-token context window and accepts high-resolution inputs (up to 4K images, any aspect ratio), enabling it to handle complex tasks such as optical character recognition, image captioning, chart and document parsing, video and scene understanding, GUI-agent workflows (e.g., interpreting screenshots, recognizing UI elements), and general vision-language reasoning. In benchmark evaluations at the 10 B-parameter scale, GLM-4.1V-9B-Thinking achieved top performance on 23 of 28 tasks.

GLM-4.5V-Flash is an open source vision-language model, designed to bring strong multimodal capabilities into a lightweight, deployable package. It supports image, video, document, and GUI inputs, enabling tasks such as scene understanding, chart and document parsing, screen reading, and multi-image analysis. Compared to larger models in the series, GLM-4.5V-Flash offers a compact footprint while retaining core VLM capabilities like visual reasoning, video understanding, GUI task handling, and complex document parsing. It can serve in “GUI agent” workflows, meaning it can interpret screenshots or desktop captures, recognize icons or UI elements, and assist with automated desktop or web-based tasks. Although it forgoes some of the largest-model performance gains, GLM-4.5V-Flash remains versatile for real-world multimodal tasks where efficiency, lower resource usage, and broad modality support are prioritized.

GLM-4.5V builds on the GLM-4.5-Air foundation, using a Mixture-of-Experts (MoE) architecture with 106 billion total parameters and 12 billion activation parameters. It achieves state-of-the-art performance among open-source VLMs of similar scale across 42 public benchmarks, excelling in image, video, document, and GUI-based tasks. It supports a broad range of multimodal capabilities, including image reasoning (scene understanding, spatial recognition, multi-image analysis), video understanding (segmentation, event recognition), complex chart and long-document parsing, GUI-agent workflows (screen reading, icon recognition, desktop automation), and precise visual grounding (e.g., locating objects and returning bounding boxes). GLM-4.5V also introduces a “Thinking Mode” switch, allowing users to choose between fast responses or deeper reasoning when needed.

GLM-4.7 is an advanced large language model designed to significantly elevate coding, reasoning, and agentic task performance. It delivers major improvements over GLM-4.6 in multilingual coding, terminal-based tasks, and real-world software engineering benchmarks such as SWE-bench and Terminal Bench. GLM-4.7 supports “thinking before acting,” enabling more stable, accurate, and controllable behavior in complex coding and agent workflows. The model also introduces strong gains in UI and frontend generation, producing cleaner webpages, better layouts, and more polished slides. Enhanced tool-using capabilities allow GLM-4.7 to perform more effectively in web browsing, automation, and agent benchmarks. Its reasoning and mathematical performance has improved substantially, showing strong results on advanced evaluation suites. GLM-4.7 is available via Z.ai, API platforms, coding agents, and local deployment for flexible adoption.

MiniMax-M2.1 is an open-source, agentic large language model designed for advanced coding, tool use, and long-horizon planning. It was released to the community to make high-performance AI agents more transparent, controllable, and accessible. The model is optimized for robustness in software engineering, instruction following, and complex multi-step workflows. MiniMax-M2.1 supports multilingual development and performs strongly across real-world coding scenarios. It is suitable for building autonomous applications that require reasoning, planning, and execution. The model weights are fully open, enabling local deployment and customization. MiniMax-M2.1 represents a major step toward democratizing top-tier agent capabilities.

Composer is Cursor’s custom-built agentic AI model optimized specifically for software engineering tasks and designed to power fast, interactive coding assistance directly within the Cursor IDE, a VS Code-derived editor enhanced with intelligent automation. It is a mixture-of-experts model trained with reinforcement learning (RL) on real-world coding problems across large codebases, so it can produce high-speed, context-aware responses, from code edits and planning to answers that understand project structure, tools, and conventions, with generation speeds roughly four times faster than similar models in benchmarks. Composer is specialized for development workflows, leveraging long-context understanding, semantic search, and limited tool access (like file editing and terminal commands) so it can solve complex engineering requests with efficient and practical outputs.

GLM-5 is Z.ai’s latest large language model built for complex systems engineering and long-horizon agentic tasks. It scales significantly beyond GLM-4.5, increasing total parameters and training data while integrating DeepSeek Sparse Attention to reduce deployment costs without sacrificing long-context capacity. The model combines enhanced pre-training with a new asynchronous reinforcement learning infrastructure called slime, improving training efficiency and post-training refinement. GLM-5 achieves best-in-class performance among open-source models across reasoning, coding, and agent benchmarks, narrowing the gap with leading frontier models. It ranks highly on evaluations such as Vending Bench 2, demonstrating strong long-term planning and operational capabilities. The model is open-sourced under the MIT License.

MiniMax M2.5 is a frontier AI model engineered for real-world productivity across coding, agentic workflows, search, and office tasks. Extensively trained with reinforcement learning in hundreds of thousands of real-world environments, it achieves state-of-the-art performance in benchmarks such as SWE-Bench Verified and BrowseComp. The model demonstrates strong architectural thinking, decomposing complex problems before generating code across more than ten programming languages. M2.5 operates at high throughput speeds of up to 100 tokens per second, enabling faster completion of multi-step tasks. It is optimized for efficient reasoning, reducing token usage and execution time compared to previous versions. With dramatically lower pricing than competing frontier models, it delivers powerful performance at minimal cost. Integrated into MiniMax Agent, M2.5 supports professional-grade office workflows, financial modeling, and autonomous task execution.

Qwen3.5 is a next-generation open-weight multimodal large language model designed to power native vision-language agents. The flagship release, Qwen3.5-397B-A17B, combines a hybrid linear attention architecture with sparse mixture-of-experts, activating only 17 billion parameters per forward pass out of 397 billion total to maximize efficiency. It delivers strong benchmark performance across reasoning, coding, multilingual understanding, visual reasoning, and agent-based tasks. The model expands language support from 119 to 201 languages and dialects while introducing a 1M-token context window in its hosted version, Qwen3.5-Plus. Built for multimodal tasks, it processes text, images, and video with advanced spatial reasoning and tool integration. Qwen3.5 also incorporates scalable reinforcement learning environments to improve general agent capabilities. Designed for developers and enterprises, it enables efficient, tool-augmented, multimodal AI workflows.

Leanstral is an open-source code agent developed by Mistral AI specifically designed to work with the Lean 4 proof assistant. The model focuses on generating code while also formally verifying its correctness against strict mathematical or software specifications. Unlike traditional coding assistants, Leanstral integrates directly with formal proof systems to ensure that generated code satisfies defined logical requirements. Its architecture is optimized for proof engineering tasks and operates efficiently with sparse model parameters. Leanstral is released under the Apache 2.0 license, making it freely accessible for developers, researchers, and organizations to use and customize. The model is designed to operate within real-world formal repositories rather than isolated problem environments. By combining code generation with formal verification, Leanstral aims to reduce the need for manual human review in complex software and mathematical development.

MiniMax M2.7 is an advanced AI model designed to enhance real-world productivity across coding, search, and office workflows. It is trained with reinforcement learning across numerous real-world environments, enabling it to handle complex, multi-step tasks effectively. The model excels in problem-solving by breaking down challenges before generating solutions across multiple programming languages. It delivers high-speed performance with rapid token generation, allowing tasks to be completed efficiently. With optimized reasoning and cost-effective pricing, it provides powerful capabilities while minimizing resource usage. It also achieves strong performance in software engineering benchmarks, reducing incident response time and improving development efficiency. Additionally, it supports advanced agentic workflows and professional-grade office tasks, making it highly versatile for modern work environments.

StarCoder and StarCoderBase are Large Language Models for Code (Code LLMs) trained on permissively licensed data from GitHub, including from 80+ programming languages, Git commits, GitHub issues, and Jupyter notebooks. Similar to LLaMA, we trained a ~15B parameter model for 1 trillion tokens. We fine-tuned StarCoderBase model for 35B Python tokens, resulting in a new model that we call StarCoder. We found that StarCoderBase outperforms existing open Code LLMs on popular programming benchmarks and matches or surpasses closed models such as code-cushman-001 from OpenAI (the original Codex model that powered early versions of GitHub Copilot). With a context length of over 8,000 tokens, the StarCoder models can process more input than any other open LLM, enabling a wide range of interesting applications. For example, by prompting the StarCoder models with a series of dialogues, we enabled them to act as a technical assistant.

The next generation of our open source large language model. This release includes model weights and starting code for pretrained and fine-tuned Llama language models — ranging from 7B to 70B parameters. Llama 2 pretrained models are trained on 2 trillion tokens, and have double the context length than Llama 1. Its fine-tuned models have been trained on over 1 million human annotations. Llama 2 outperforms other open source language models on many external benchmarks, including reasoning, coding, proficiency, and knowledge tests. Llama 2 was pretrained on publicly available online data sources. The fine-tuned model, Llama-2-chat, leverages publicly available instruction datasets and over 1 million human annotations. We have a broad range of supporters around the world who believe in our open approach to today’s AI — companies that have given early feedback and are excited to build with Llama 2.

Code Llama is a large language model (LLM) that can use text prompts to generate code. Code Llama is state-of-the-art for publicly available LLMs on code tasks, and has the potential to make workflows faster and more efficient for current developers and lower the barrier to entry for people who are learning to code. Code Llama has the potential to be used as a productivity and educational tool to help programmers write more robust, well-documented software. Code Llama is a state-of-the-art LLM capable of generating code, and natural language about code, from both code and natural language prompts. Code Llama is free for research and commercial use. Code Llama is built on top of Llama 2 and is available in three models: Code Llama, the foundational code model; Codel Llama - Python specialized for Python; and Code Llama - Instruct, which is fine-tuned for understanding natural language instructions.

Yi-Large is a proprietary large language model developed by 01.AI, offering a 32k context length with both input and output costs at $2 per million tokens. It stands out with its advanced capabilities in natural language processing, common-sense reasoning, and multilingual support, performing on par with leading models like GPT-4 and Claude3 in various benchmarks. Yi-Large is designed for tasks requiring complex inference, prediction, and language understanding, making it suitable for applications like knowledge search, data classification, and creating human-like chatbots. Its architecture is based on a decoder-only transformer with enhancements such as pre-normalization and Group Query Attention, and it has been trained on a vast, high-quality multilingual dataset. This model's versatility and cost-efficiency make it a strong contender in the AI market, particularly for enterprises aiming to deploy AI solutions globally.

Composer 1.5 is the latest agentic coding model from Cursor that balances speed and intelligence for everyday code tasks by scaling reinforcement learning approximately 20x more than its predecessor, enabling stronger performance on real-world programming challenges. It’s designed as a “thinking model” that generates internal reasoning tokens to analyze a user’s codebase and plan next steps, responding quickly to simple problems and engaging deeper reasoning on complex ones, while remaining interactive and fast for daily development workflows. To handle long-running tasks, Composer 1.5 introduces self-summarization, allowing the model to compress and carry forward context when it reaches context limits, which helps maintain accuracy across varying input lengths. Internal benchmarks show it surpasses Composer 1 in coding tasks, especially on more difficult issues, making it more capable for interactive use within Cursor’s environment.

PaLM 2 is our next generation large language model that builds on Google’s legacy of breakthrough research in machine learning and responsible AI. It excels at advanced reasoning tasks, including code and math, classification and question answering, translation and multilingual proficiency, and natural language generation better than our previous state-of-the-art LLMs, including PaLM. It can accomplish these tasks because of the way it was built – bringing together compute-optimal scaling, an improved dataset mixture, and model architecture improvements. PaLM 2 is grounded in Google’s approach to building and deploying AI responsibly. It was evaluated rigorously for its potential harms and biases, capabilities and downstream uses in research and in-product applications. It’s being used in other state-of-the-art models, like Med-PaLM 2 and Sec-PaLM, and is powering generative AI features and tools at Google, like Bard and the PaLM API.