Top Devstral 2 Alternatives

Kimi K3.1 Pro is a multimodal large language model developed by Moonshot AI to support advanced reasoning, autonomous AI agents, software development, 3D creation, and complex knowledge work. The model is based on a mixture-of-experts architecture containing approximately 2.8 trillion total parameters and 104 billion active parameters. Its one-million-token context window enables the processing of extensive documents, source code repositories, technical materials, and long-running task histories. Kimi K3.1 Pro builds on the Kimi K3 foundation with additional reinforcement learning and post-training optimization across programming, analytical reasoning, and agentic tasks. Its software engineering capabilities include code generation, debugging, terminal operations, automated testing, application development, and multistep problem-solving across existing repositories. The model can develop interactive websites, animated interfaces, browser-based 3D environments, and game projects using tools and frameworks such as Three.js, Unity, Unreal Engine, and Godot. Its multimodal reasoning capabilities allow it to interpret images, sketches, and visual references when generating technical designs or structured outputs. Three-dimensional modeling features include CAD script generation, parametric geometry creation, asset modification, and integration with tools such as Blender. Financial reasoning capabilities support complex spreadsheet models, financial statement analysis, investment research, cross-document verification, and evidence-based reporting. Kimi K3.1 Pro also supports tool-assisted workflows in which AI agents plan tasks, execute actions, evaluate results, and refine outputs across multiple steps. The model is available through Kimi's web and mobile applications, Kimi Work, Kimi Code, and the Kimi API, with a Highspeed variant providing faster inference and an open-weight release planned for November 4, 2026.

Amp is a frontier coding agent designed to redefine how developers interact with AI during software development. Built for use in terminals and modern editors, Amp allows engineers to orchestrate powerful AI agents that can reason across entire repositories, not just isolated files. It supports advanced workflows such as large-scale refactors, architecture exploration, agent-generated code reviews, and parallel course correction with forced tool usage. Amp integrates leading AI models and layers them with robust context management, subagents, and continuous tooling improvements. Developers can let agents run autonomously, trusting them to produce consistent, high-quality results across complex projects. With strong community adoption, rapid feature releases, and a focus on real engineering use cases, Amp stands out as a premium, agent-first coding platform. It empowers developers to ship faster, explore deeper, and build systems that would otherwise require significantly more time and effort.

NVIDIA Nemotron 3.5 Super VL is an open multimodal artificial intelligence model designed to support specialized autonomous agents, enterprise automation, and complex reasoning tasks. Introduced in October 2026, the model combines language processing, visual understanding, software engineering capabilities, and external tool interactions within a unified architecture. It uses a 120-billion-parameter mixture-of-experts design with approximately 12 billion active parameters per token to optimize inference efficiency. The model was distilled from NVIDIA Nemotron 3 Ultra and incorporates additional training focused on coding, multimodal reasoning, and agentic task execution. Its vision-language capabilities support analysis of images, charts, tables, diagrams, multipage documents, and video content. Document intelligence workflows can extract structured information, interpret visual evidence, connect information across pages, and generate summaries or reports. Software engineering features enable the model to analyze repositories, identify programming errors, test code, and recommend modifications across development workflows. Its agentic capabilities support web research, external tool usage, multistep problem-solving, cybersecurity vulnerability analysis, and enterprise process automation. NVIDIA NeMo provides customization tools for supervised fine-tuning, low-rank adaptation, and reinforcement learning using proprietary datasets, organizational policies, and specialized workflows. The model incorporates LiLiCorr speculative decoding, multi-token prediction, context reuse, and NVFP4 quantization to improve inference performance on compatible hardware. Nemotron 3.5 Super VL is distributed with open weights and supports local, cloud, and data center deployment through NVIDIA infrastructure and tools such as NVIDIA NIM.

Muse Spark Open is a large-scale open-weight artificial intelligence model developed by Meta for multimodal understanding, software engineering, autonomous agents, and long-horizon reasoning. Released in October 2026, the model provides downloadable weights that allow developers to deploy and customize it outside Meta's hosted AI services. It uses a mixture-of-experts architecture containing 816 billion total parameters, with approximately 42 billion parameters activated during inference. Its one-million-token context window supports processing extensive documents, large software repositories, and lengthy sequences of instructions and interactions. The model is designed for complex coding workflows involving source code analysis, debugging, feature implementation, application development, and software testing. Its multimodal capabilities enable AI agents to interpret visual information and combine that understanding with reasoning and automated computer interactions. Meta demonstrated the model performing local software development tasks and controlling DaVinci Resolve to analyze video footage and prepare edited content. Muse Spark Open supports on-device deployment using two Apple M5 Ultra Mac Studio computers running the MLX framework, with additional compatibility across AMD, Intel, and NVIDIA hardware. Supported inference frameworks include SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp, and Unsloth, with optimized GPU kernels for improved processing performance. Developers can customize the model using PyTorch and TorchTitan to adapt its behavior to specialized tasks, organizational requirements, and application workflows. Muse Spark Open is distributed under the Muse Model License with limited commercial terms and is available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry Models, and several hosted API providers.

Laguna S 2.1 represents a state-of-the-art open weight coding model that focuses on the completion of long-term projects and demonstrates exceptional reasoning abilities. With a Mixture-of-Experts architecture comprising 118 billion parameters, it engages 8 billion parameters per token and supports a context window of up to one million tokens in both cognitive and non-cognitive modes. The model’s optimized active size enables it to execute complex tasks on local systems while remaining competitive with much larger models across a variety of benchmarks, such as terminal usage, software development, codebase question answering, and tool application. Built for durability, Laguna S 2.1 is adept at addressing demanding challenges with an emphasis on thorough verification and a willingness to backtrack when necessary, rather than hastily claiming victory. In real-world scenarios, it has successfully engineered a browser rendering engine from the ground up, improved an agent harness for faster execution and lower memory requirements, and conducted comprehensive mathematical investigations using the tools available in its environment, showcasing its adaptability and proficiency. This remarkable array of capabilities positions Laguna S 2.1 as an invaluable asset for developers in search of cutting-edge solutions, making it a top choice in the ever-evolving landscape of coding models.

NVIDIA Nemotron 3.5 Bolt is a family of open artificial intelligence models developed to provide efficient reasoning, autonomous agent capabilities, and conversational intelligence on local computing hardware. Introduced in October 2026, the series includes Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. The 74B-A9B variant uses a mixture-of-experts architecture containing 74 billion total parameters, with approximately 9 billion parameters activated during inference. Its native 2-bit quantization enables deployment within approximately 24GB of memory while preserving resources for other applications. The model is designed for agentic workflows that require reasoning, instruction following, tool interactions, and execution of multiple connected tasks. Developers can use Bolt 74B-A9B with local agent frameworks to conduct web research, generate software, analyze code, investigate programming errors, and perform debugging activities. NVIDIA reports strong results on agent benchmarks including PinchBench and BrowseComp, which evaluate autonomous task completion and information retrieval capabilities. The smaller Bolt E4B variant is designed for low-latency conversational AI, supporting local voice assistants, interactive game characters, and embedded interfaces. Its 4GB memory footprint makes it suitable for compatible GeForce RTX, NVIDIA Jetson, and DGX Spark systems that require efficient local inference. Running these models locally allows organizations to retain greater control over sensitive application data, reduce dependence on cloud inference, and integrate AI capabilities into existing software environments. Nemotron 3.5 Bolt 74B-A9B is available through Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth, and vLLM, while Bolt E4B is scheduled for release on October 23, 2026.

GLM-5.3-Flash is an efficiency-focused multimodal AI model from Z.ai that combines advanced reasoning, coding, agentic execution, and visual intelligence. It is the first GLM-5-series model designed with native multimodal capabilities, allowing it to work directly with both textual and visual inputs. The architecture uses 320 billion total parameters while activating only 18 billion at a time, significantly reducing the amount of computation required for inference. A hybrid attention design blends linear attention for local information with sparse attention for retrieving important context from much larger inputs. Z.ai also uses technologies such as IndexPool and Manifold-Constrained Hyper-Connections to improve memory efficiency, latency, and model scaling. The model can operate with context windows of up to one million tokens, making it suitable for large repositories, lengthy documents, extended agent sessions, and complex multimodal workflows. GLM-5.3-Flash was trained on a 30-trillion-token multimodal corpus intended to strengthen reasoning across code, images, interfaces, documents, spreadsheets, presentations, and other business artifacts. In software development scenarios, the model can visually inspect rendered applications, evaluate its own output, and iteratively correct layout, functionality, or interaction issues. Z.ai’s reported benchmark results show large improvements over GLM-5.2 in areas such as software engineering and automation, while placing GLM-5.3-Flash close to leading frontier systems on several coding and agentic evaluations. Before its formal release, the model was anonymously tested under the name ox-alpha on OpenCode and OpenRouter, where Z.ai says it became one of the most widely used models during its testing period. GLM-5.3-Flash is available through Z.ai’s API and coding products as well as through downloadable weights on Hugging Face, with deployment support for SGLang, vLLM, and TokenSpeed.

DeepSeek-V4-Pro is a next-generation Mixture-of-Experts language model designed to deliver high performance across reasoning, coding, and long-context AI tasks. It features a massive architecture with 1.6 trillion total parameters and 49 billion activated parameters, enabling efficient computation while maintaining strong capabilities. The model supports an industry-leading context window of up to one million tokens, allowing it to process extremely large datasets, documents, and workflows. Its hybrid attention mechanism combines advanced techniques to optimize long-context efficiency and reduce computational requirements. DeepSeek-V4-Pro is trained on over 32 trillion tokens, enhancing its knowledge base and reasoning abilities. It incorporates advanced optimization methods to improve training stability and convergence. The model supports multiple reasoning modes, including fast responses and deep analytical thinking for complex problem solving. It performs strongly across benchmarks in coding, mathematics, and knowledge-based tasks. The architecture is designed for agentic workflows, enabling it to handle multi-step tasks and tool-based interactions. As an open-source model, it offers flexibility for customization and deployment across various environments. It also supports efficient memory usage and reduced inference costs compared to previous versions. The model’s capabilities make it suitable for both research and enterprise applications. Overall, DeepSeek-V4-Pro represents a significant advancement in scalable, high-performance AI with long-context intelligence.

Laguna M.1 is recognized as Poolside's premier model for agentic coding, meticulously designed in-house to optimize software development processes. This sophisticated model incorporates 225 billion parameters and employs a Mixture of Experts architecture with 23 billion parameters activated, all trained on a colossal dataset of 30 trillion tokens using a network of 6,144 NVIDIA H200 GPUs. Poolside committed to developing Laguna M.1 from the ground up, utilizing proprietary data, a specialized training codebase, and an asynchronous on-policy reinforcement learning strategy within its agent framework, all specifically oriented towards agentic coding applications. The model's architecture is crafted to deliver top-tier performance within Poolside's coding agent, empowering it to adeptly reason through programming tasks, engage with an array of tools, modify code, run tests, and support extensive autonomous development sessions. Tailored for developers and teams facing complex coding obstacles, Laguna M.1 boasts enhanced capabilities in reasoning, understanding architecture, managing terminal actions, and executing multi-step processes, far exceeding the abilities of lighter models. Overall, its comprehensive feature set establishes it as an indispensable tool for professionals immersed in high-stakes software projects, making it a vital component in the landscape of agentic coding solutions.

Inkling is an open-weights multimodal AI model from Thinking Machines built to support customization, agentic workflows, coding, reasoning, vision, audio, and enterprise AI use cases. The model is a Mixture-of-Experts transformer with 975 billion total parameters, 41 billion active parameters, 256 routed experts per MoE layer, and six routed experts active per token. It supports context windows up to 1 million tokens and was pretrained on 45 trillion tokens across text, images, audio, and video. Inkling is designed as a broad foundation model rather than a narrowly optimized benchmark model, giving it balanced capabilities across reasoning, coding, factuality, instruction following, vision, audio, tool use, and safety. Its controllable thinking effort lets developers adjust how much computation and generated reasoning the model uses, helping teams balance quality, latency, and cost for different production needs. The model can run agentic coding tasks, use tools, create web apps, generate polished multi-page artifacts, reason over long contexts, and work through iterative refinement loops. For multimodal tasks, Inkling can process images, answer questions about visual content, transcribe and reason over audio, follow spoken instructions, and combine visual reasoning with code-based tools such as Python. Thinking Machines trained Inkling for calibration, instruction following, factual reliability, refusal behavior, and safety across multiple modalities, including evaluations for dangerous capabilities and human-AI threat vectors. Inkling is available on Tinker for fine-tuning, with 64K and 256K context options, an Inkling Playground for testing, cookbook recipes, and support for multimodal post-training workflows. Its full weights are available on Hugging Face, and deployment support is available through APIs and infrastructure partners such as TogetherAI, Fireworks, Modal, Databricks, Baseten, SGLang, vLLM, llama.cpp, and transformers.

DeepCoder, a fully open-source initiative for code reasoning and generation, has been created through a collaboration between the Agentica Project and Together AI. Built on the foundation of DeepSeek-R1-Distilled-Qwen-14B, it has been fine-tuned using distributed reinforcement learning techniques, achieving an impressive accuracy of 60.6% on LiveCodeBench, which represents an 8% improvement compared to its predecessor. This remarkable performance positions it competitively alongside proprietary models such as o3-mini (2025-01-031 Low) and o1, all while operating with a streamlined 14 billion parameters. The training process was intensive, lasting 2.5 weeks on a fleet of 32 H100 GPUs and utilizing a meticulously curated dataset comprising around 24,000 coding challenges obtained from reliable sources such as TACO-Verified, PrimeIntellect SYNTHETIC-1, and submissions to LiveCodeBench. Each coding challenge was required to include a valid solution paired with at least five unit tests to ensure robustness during the reinforcement learning phase. Additionally, DeepCoder employs innovative methods like iterative context lengthening and overlong filtering to effectively handle long-range contextual dependencies, allowing it to tackle complex coding tasks with proficiency. This distinctive approach not only enhances DeepCoder's accuracy and reliability in code generation but also positions it as a significant player in the landscape of code generation models. As a result, developers can rely on its capabilities for diverse programming challenges.

Devstral Small 2 is a condensed, 24 billion-parameter variant of Mistral AI's groundbreaking coding-focused models, made available under the adaptable Apache 2.0 license to support both local use and API access. Alongside its more extensive sibling, Devstral 2, it offers "agentic coding" capabilities tailored for low-computational environments, featuring a substantial 256K-token context window that enables it to understand and alter entire codebases with ease. With a performance score nearing 68.0% on the widely recognized SWE-Bench Verified code-generation benchmark, Devstral Small 2 distinguishes itself within the realm of open-weight models that are much larger. Its compact structure and efficient design allow it to function effectively on a single GPU or even in CPU-only setups, making it an excellent option for developers, small teams, or hobbyists who may lack access to extensive data-center facilities. Moreover, despite being smaller, Devstral Small 2 retains critical functionalities found in its larger counterparts, such as the capability to reason through multiple files and adeptly manage dependencies, ensuring that users enjoy substantial coding support. This combination of efficiency and high performance positions it as an indispensable asset for the coding community. Additionally, its user-friendly approach ensures that both novice and experienced programmers can leverage its capabilities without significant barriers.

Devstral represents a joint initiative by Mistral AI and All Hands AI, creating an open-source large language model designed explicitly for the field of software engineering. This innovative model exhibits exceptional skill in navigating complex codebases, efficiently managing edits across multiple files, and tackling real-world issues, achieving an impressive 46.8% score on the SWE-Bench Verified benchmark, which positions it ahead of all other open-source models. Built upon the foundation of Mistral-Small-3.1, Devstral features a vast context window that accommodates up to 128,000 tokens. It is optimized for peak performance on advanced hardware configurations, such as Macs with 32GB of RAM or Nvidia RTX 4090 GPUs, and is compatible with several inference frameworks, including vLLM, Transformers, and Ollama. Released under the Apache 2.0 license, Devstral is readily available on various platforms, including Hugging Face, Ollama, Kaggle, Unsloth, and LM Studio, enabling developers to effortlessly incorporate its features into their applications. This model not only boosts efficiency for software engineers but also acts as a crucial tool for anyone engaged in coding tasks, thereby broadening its utility and appeal across the tech community. Furthermore, its open-source nature encourages continuous improvement and collaboration among developers worldwide.

DeepSWE represents a groundbreaking advancement in open-source coding agents, harnessing the Qwen3-32B foundation model trained exclusively through reinforcement learning (RL) without the aid of supervised fine-tuning or proprietary model distillation. Developed using rLLM, which is Agentica's open-source RL framework tailored for language-driven agents, DeepSWE functions effectively within a simulated development environment provided by the R2E-Gym framework. This setup equips it with a range of tools, such as a file editor, search functions, shell execution, and submission capabilities, allowing the agent to adeptly navigate extensive codebases, modify multiple files, compile code, execute tests, and iteratively generate patches or fulfill intricate engineering tasks. In addition to mere code generation, DeepSWE exhibits sophisticated emergent behaviors; when confronted with bugs or feature requests, it engages in critical reasoning regarding edge cases, searches for existing tests in the codebase, proposes patches, creates additional tests to avert regressions, and adapts its cognitive strategies based on the specific challenges presented. This remarkable adaptability and efficiency position DeepSWE as a formidable asset in the software development landscape, empowering developers to tackle complex projects with greater ease and confidence. Its ability to learn from each interaction further enhances its performance, ensuring continuous improvement over time.

The Laguna XS 2.1 represents a sophisticated advancement in coding models, functioning as an open weight agentic system that excels in executing long-duration tasks on local machines. It boasts a robust 33-billion-parameter Mixture-of-Experts architecture, activating 3 billion parameters per token, while preserving the efficient design of its predecessor, Laguna XS.2, and significantly enhancing its capabilities in multilingual software engineering and terminal-related tasks. This model is meticulously crafted to support coding agents in reviewing code repositories, navigating complex changes, leveraging diverse tools, executing commands, and ensuring seamless progress throughout extensive projects. With an impressive context window of 256K, it empowers agents to adeptly handle large codebases, maintain extensive histories, and navigate intricate multi-step workflows. The Laguna XS 2.1 also enjoys compatibility with various platforms like vLLM, SGLang, NVIDIA TensorRT-LLM, Hugging Face Transformers, and Ollama, with aspirations for future native support from llama.cpp. Offered in multiple checkpoint formats such as BF16, FP8, INT4, and NVFP4, it allows developers to choose between high fidelity and configurations designed for environments with restricted VRAM or processing capacity. This versatility not only enhances its usability across different development frameworks but also positions it as a prime choice for diverse programming needs and settings. Furthermore, its ability to adapt to varying project demands makes it a valuable asset for developers seeking efficiency and performance in their workflows.

Mistral Small 4 is a powerful open-source AI model introduced by Mistral AI to deliver advanced reasoning, multimodal understanding, and coding capabilities in a single system. The model represents the latest evolution in the Mistral Small family and consolidates multiple specialized AI technologies into one unified architecture. It integrates the reasoning capabilities of Magistral, the multimodal functionality of Pixtral, and the coding intelligence of Devstral. This design allows the model to handle tasks ranging from conversational assistance and research analysis to software development and visual data processing. Mistral Small 4 supports both text and image inputs, enabling applications such as document parsing, visual analysis, and interactive AI systems. Its mixture-of-experts architecture includes 128 experts with a small subset activated per token, allowing efficient resource usage while maintaining strong performance. The model also introduces a configurable reasoning effort parameter that allows developers to control the balance between speed and analytical depth. A large 256k context window enables it to process lengthy conversations, documents, and complex reasoning workflows. Performance optimizations significantly reduce latency and increase throughput compared with previous versions of the model. The system is designed for deployment across various environments, including cloud infrastructure, enterprise systems, and research environments. Developers can access the model through platforms such as Hugging Face, Transformers, and optimized inference frameworks. Released under the Apache 2.0 open-source license, Mistral Small 4 allows organizations to customize, fine-tune, and deploy AI solutions tailored to their specific needs. By combining reasoning, multimodal processing, and coding intelligence in one model, Mistral Small 4 simplifies AI integration for modern applications.

Molmo 2 introduces a state-of-the-art collection of open vision-language models, offering fully accessible weights, training data, and code, which enhances the capabilities of the original Molmo series by extending grounded image comprehension to include video and various image inputs. This significant upgrade facilitates advanced video analysis tasks such as pointing, tracking, dense captioning, and question-answering, all exhibiting strong spatial and temporal reasoning across multiple frames. The suite is comprised of three unique models: an 8 billion-parameter version designed for thorough video grounding and QA tasks, a 4 billion-parameter model that emphasizes efficiency, and a 7 billion-parameter model powered by Olmo, featuring a completely open end-to-end architecture that integrates the core language model. Remarkably, these latest models outperform their predecessors on important benchmarks, establishing new benchmarks for open-model capabilities in image and video comprehension tasks. Additionally, they frequently compete with much larger proprietary systems while being trained on a significantly smaller dataset compared to similar closed models, illustrating their impressive efficiency and performance in the domain. This noteworthy accomplishment signifies a major step forward in making AI-driven visual understanding technologies more accessible and effective, paving the way for further innovations in the field. The advancements presented by Molmo 2 not only enhance user experience but also broaden the potential applications of AI in various industries.

Recursive Flint is a state-of-the-art language model boasting 40 billion parameters, crafted as a mixture-of-experts particularly for local coding tasks. It utilizes approximately 2.5 billion parameters for every token it processes, allowing for advanced reasoning, multi-turn tool usage, and support for context windows reaching up to 512,000 tokens. Developers have the opportunity to integrate Flint into coding agents, enabling them to inspect codebases, fix bugs, create new features, and formulate tests while running inference directly on their own systems. The model's weights are accessible under the Apache 2.0 license, fostering broader accessibility for users. Flint supports a range of platforms, including llama.cpp, Ollama, Unsloth, LM Studio, vLLM, and SGLang. In addition, Recursive provides rec, an inference engine finely tuned for optimal performance on Apple Silicon M4 Pro and M5 Pro/Max chips. To facilitate the implementation of the model on local machines, detailed setup documentation and practical examples are available for developers. This versatile tool not only aids developers and technical teams in running a code-focused model tailored to their individual system needs but also enhances their overall coding efficiency. As a result, users can significantly boost their development workflows and achieve their programming goals with greater ease and speed than ever before.

Pixtral Large is a comprehensive multimodal model developed by Mistral AI, boasting an impressive 124 billion parameters that build upon their earlier Mistral Large 2 framework. The architecture consists of a 123-billion-parameter multimodal decoder paired with a 1-billion-parameter vision encoder, which empowers the model to adeptly interpret diverse content such as documents, graphs, and natural images while maintaining excellent text understanding. Furthermore, Pixtral Large can accommodate a substantial context window of 128,000 tokens, enabling it to process at least 30 high-definition images simultaneously with impressive efficiency. Its performance has been validated through exceptional results in benchmarks like MathVista, DocVQA, and VQAv2, surpassing competitors like GPT-4o and Gemini-1.5 Pro. The model is made available for research and educational use under the Mistral Research License, while also offering a separate Mistral Commercial License for businesses. This dual licensing approach enhances its appeal, making Pixtral Large not only a powerful asset for academic research but also a significant contributor to advancements in commercial applications. As a result, the model stands out as a multifaceted tool capable of driving innovation across various fields.

GLM-4.1V represents a cutting-edge vision-language model that provides a powerful and efficient multimodal ability for interpreting and reasoning through different types of media, such as images, text, and documents. The 9-billion-parameter variant, referred to as GLM-4.1V-9B-Thinking, is built on the GLM-4-9B foundation and has been refined using a distinctive training method called Reinforcement Learning with Curriculum Sampling (RLCS). With a context window that accommodates 64k tokens, this model can handle high-resolution inputs, supporting images with a resolution of up to 4K and any aspect ratio, enabling it to perform complex tasks like optical character recognition, image captioning, chart and document parsing, video analysis, scene understanding, and GUI-agent workflows, which include interpreting screenshots and identifying UI components. In benchmark evaluations at the 10 B-parameter scale, GLM-4.1V-9B-Thinking achieved remarkable results, securing the top performance in 23 of the 28 tasks assessed. These advancements mark a significant progression in the fusion of visual and textual information, establishing a new benchmark for multimodal models across a variety of applications, and indicating the potential for future innovations in this field. This model not only enhances existing workflows but also opens up new possibilities for applications in diverse domains.

SWE-2 Flash is an AI coding and software engineering model developed by Cognition to deliver high-speed programming capabilities with lower inference costs than larger frontier models. The model uses a 276-billion-parameter mixture-of-experts architecture with approximately 12 billion active parameters, based on Thinking Machines' Inkling-Small foundation model. Cognition enhances the underlying model through a dedicated mid-training process followed by large-scale reinforcement learning focused on software engineering and general reasoning tasks. Its training incorporates several trillion tokens of carefully selected technical material, including source code, scientific publications, articles, and other datasets relevant to complex problem-solving. SWE-2 Flash supports automated code generation, software maintenance, code reviews, feature development, documentation updates, and other tasks performed by autonomous development agents. The model is optimized for efficient inference and generates approximately 200 tokens per second, compared with approximately 60 tokens per second for the larger SWE-2 model. Cognition reports that SWE-2 Flash operates at one-fifth the cost of SWE-2 while maintaining competitive performance across several software engineering evaluations. Its published benchmark results include 24.0% on FrontierCode 2.0, 45.5% on FrontierCode 1.1 Main, 60.1% on DeepSWE 1.1, and 89.9% on SWE-Bench Pro v2 Hard. The model also demonstrates cybersecurity-related capabilities, achieving 34.8% on ExploitBench compared with 11.3% for its original foundation model. Developers can access SWE-2 Flash through Devin Desktop and CLI for interactive programming assistance and agent-driven engineering workflows. Cognition is developing a secure program for sharing model weights with approved industry partners and research organizations, potentially enabling fine-tuning and deployment on private or on-premises infrastructure.

Muse Glimmer is a cutting-edge model boasting 30 billion parameters, crafted by Meta Superintelligence Labs, specifically optimized for seamless local agent functionality. Its streamlined architecture enables operation on standard Mac or PC systems with a single consumer GPU, making it suitable for a range of applications, including local agent management, programming tasks, function invocation, and evaluations within LLM-as-a-judge scenarios, all without needing cloud services or an internet connection. This groundbreaking model features sophisticated abilities like long-horizon execution, precise tool invocation, multimodal understanding, expanded memory for contextual awareness, and proficient instruction adherence. It excels in performing comprehensive tasks as an agent, adeptly navigates complex multi-step reasoning across extensive workflows, and can recover effectively from unexpected tool interactions. Additionally, it interprets interleaved text and images through a specialized perception encoder tailored for analyzing screenshots, graphs, and various document types. Beyond its primary functions, Muse Glimmer is designed to work harmoniously with OpenClaw and other orchestration frameworks, allowing for customizable reasoning capabilities and has been trained on a rich dataset that spans over 100 languages. The adaptability of this model not only enhances its effectiveness across different fields but also positions it as a significant asset in the evolving landscape of AI applications. Its innovative features and user-friendly deployment make it a versatile choice for professionals seeking to leverage AI for complex problem-solving.

Composer is an AI model developed by Cursor, specifically designed for software engineering tasks, providing fast and interactive coding assistance within the Cursor IDE, an upgraded version of a VS Code-based editor that features intelligent automation capabilities. This model uses a mixture-of-experts framework and reinforcement learning (RL) to address real-world coding challenges encountered in large codebases, allowing it to offer quick, contextually relevant responses that include code adjustments, planning, and insights into project frameworks, tools, and conventions, achieving generation speeds that are nearly four times faster than those of its peers in performance evaluations. With a focus on the development workflow, Composer incorporates long-context understanding, semantic search functionalities, and limited tool access (including file manipulation and terminal commands) to effectively resolve complex engineering questions with practical and efficient solutions. Its distinctive architecture not only enables adaptability across various programming environments but also ensures that users receive personalized support tailored to their individual coding requirements. Furthermore, the versatility of Composer allows it to evolve alongside the ever-changing landscape of software development, making it an invaluable resource for developers seeking to enhance their coding experience.

Reka Flash 3 stands as a state-of-the-art multimodal AI model, boasting 21 billion parameters and developed by Reka AI, to excel in diverse tasks such as engaging in general conversations, coding, adhering to instructions, and executing various functions. This innovative model skillfully processes and interprets a wide range of inputs, which includes text, images, video, and audio, making it a compact yet versatile solution fit for numerous applications. Constructed from the ground up, Reka Flash 3 was trained on a diverse collection of datasets that include both publicly accessible and synthetic data, undergoing a thorough instruction tuning process with carefully selected high-quality information to refine its performance. The concluding stage of its training leveraged reinforcement learning techniques, specifically the REINFORCE Leave One-Out (RLOO) method, which integrated both model-driven and rule-oriented rewards to enhance its reasoning capabilities significantly. With a remarkable context length of 32,000 tokens, Reka Flash 3 effectively competes against proprietary models such as OpenAI's o1-mini, making it highly suitable for applications that demand low latency or on-device processing. Operating at full precision, the model requires a memory footprint of 39GB (fp16), but this can be optimized down to just 11GB through 4-bit quantization, showcasing its flexibility across various deployment environments. Furthermore, Reka Flash 3's advanced features ensure that it can adapt to a wide array of user requirements, thereby reinforcing its position as a leader in the realm of multimodal AI technology. This advancement not only highlights the progress made in AI but also opens doors to new possibilities for innovation across different sectors.

Aion 1.0 Plan is a groundbreaking local agentic reasoning framework developed by Microsoft for Windows, enabling comprehensive agentic workflows on devices without dependence on cloud services or additional per-token costs. Featuring an impressive architecture with 14 billion parameters and a context length of 32K, this model is seamlessly integrated into Windows on compatible hardware. Unlike smaller on-device models that simply focus on basic text processing, Aion 1.0 Plan is crafted for sophisticated local agentic reasoning, empowering applications to grasp user intentions, utilize various tools, handle file management, and coordinate sub-agents on the device autonomously. This framework marks a significant advancement in Microsoft's lineup of on-device small language models, designed for effective local execution and indicating a transition from scalable text intelligence to more refined local planning capabilities. Aion 1.0 Plan plays a vital role in the broader initiative of Windows to provide “unmetered intelligence,” wherein advanced models address intricate challenges while local counterparts ensure continuous, affordable agent workflows. This evolution not only enhances user-device interactions but also significantly boosts productivity and simplifies everyday computing tasks, representing a major step towards more intuitive technology. As such, users can expect a more tailored experience that aligns closely with their individual needs and working styles.

LongCat-2.0 signifies a remarkable leap forward in the field of language models, boasting an impressive 1.6 trillion parameters through a Mixture-of-Experts architecture that utilizes AI ASIC superpods, with around 48 billion parameters activated per token, demonstrating outstanding proficiency in coding and agentic functions. This model notably surpasses its predecessors by incorporating a large-scale sparse architecture along with specialized post-training techniques designed specifically for applications in real-world software development, tool usage, long-context reasoning, and intricate agent operations. Entirely built and executed on AI ASIC superpods, LongCat-2.0's pretraining involved processing over 35 trillion tokens and countless accelerator hours, highlighting the forefront of training techniques on state-of-the-art hardware. To further enhance its capabilities on tasks that require long-term contextual awareness, the model integrates LongCat Sparse Attention and is trained with hundreds of billions of tokens derived from 1M-context datasets, which empowers it to adeptly handle ultra-long context challenges and maintain a comprehensive understanding of extensive documents. This unique blend of features not only establishes LongCat-2.0 as an innovative leader in advanced language models but also sets a new benchmark for future developments in the domain. Its capabilities are likely to inspire a new wave of research and applications in the field.

DeepScaleR is an advanced language model featuring 1.5 billion parameters, developed from DeepSeek-R1-Distilled-Qwen-1.5B through a unique blend of distributed reinforcement learning and a novel technique that gradually increases its context window from 8,000 to 24,000 tokens throughout training. The model was constructed using around 40,000 carefully curated mathematical problems taken from prestigious competition datasets, such as AIME (1984–2023), AMC (pre-2023), Omni-MATH, and STILL. With an impressive accuracy rate of 43.1% on the AIME 2024 exam, DeepScaleR exhibits a remarkable improvement of approximately 14.3 percentage points over its base version, surpassing even the significantly larger proprietary O1-Preview model. Furthermore, its outstanding performance on various mathematical benchmarks, including MATH-500, AMC 2023, Minerva Math, and OlympiadBench, illustrates that smaller, finely-tuned models enhanced by reinforcement learning can compete with or exceed the performance of larger counterparts in complex reasoning challenges. This breakthrough highlights the promising potential of streamlined modeling techniques in advancing mathematical problem-solving capabilities, encouraging further exploration in the field. Moreover, it opens doors for developing more efficient models that can tackle increasingly challenging problems with great efficacy.

Ling 3.0 Flash is an evolved language model specifically designed for long-term agent tasks, featuring rapid response capabilities, low activation levels, and reliable tool utilization. With a Mixture-of-Experts architecture, it encompasses an impressive total of 124 billion parameters, activating 5.1 billion parameters for each token, which optimizes its performance while ensuring efficient inference. The model showcases a remarkable native context window of 256K tokens, expandable to a maximum of 1 million tokens, facilitating effective information retrieval from extensive contexts. In comparison to its earlier version, the original Flash model, Ling 3.0 Flash offers superior stability for extended operations, enhances tool-calling accuracy, adheres more closely to instructions, and shows improved compatibility with agent harnesses and coding tasks. Furthermore, its advanced spatial awareness capabilities allow it to construct grids of physical scenes and assess relative positions with precision, while its hybrid reasoning abilities increase success rates across various task complexities. This model not only represents a substantial advancement in language modeling technology but also ensures users can attain exceptional performance across a wide array of applications, thus broadening its potential use cases. Overall, Ling 3.0 Flash stands out as a groundbreaking development in the field, likely to influence future applications significantly.

DeepSeek-V4 is a cutting-edge open-source AI model built to deliver exceptional performance in reasoning, coding, and large-scale data processing. It supports an industry-leading one million token context window, allowing it to manage long documents and complex tasks efficiently. The model includes two variants: DeepSeek-V4-Pro, which offers 1.6 trillion parameters with 49 billion active for top-tier performance, and DeepSeek-V4-Flash, which provides a faster and more cost-effective alternative. DeepSeek-V4 introduces structural innovations such as token-wise compression and sparse attention, significantly reducing computational overhead while maintaining accuracy. It is designed with strong agentic capabilities, enabling seamless integration with AI agents and multi-step workflows. The model excels in domains such as mathematics, coding, and scientific reasoning, outperforming many open-source alternatives. It also supports flexible reasoning modes, allowing users to optimize for speed or depth depending on the task. DeepSeek-V4 is compatible with popular APIs, making it easy to integrate into existing systems. Its open-source nature allows developers to customize and scale it according to their needs. The model is already being used in advanced coding agents and automation workflows. It delivers a strong balance of performance, efficiency, and scalability for real-world applications. Overall, DeepSeek-V4 represents a major advancement in accessible, high-performance AI technology.

This repository presents the research preview for LongLLaMA, an innovative large language model capable of handling extensive contexts, reaching up to 256,000 tokens or potentially even more. Built on the OpenLLaMA framework, LongLLaMA has been fine-tuned using the Focused Transformer (FoT) methodology. The foundational code for this model comes from Code Llama. We are excited to introduce a smaller 3B base version of the LongLLaMA model, which is not instruction-tuned, and it will be released under an open license (Apache 2.0). Accompanying this release is inference code that supports longer contexts, available on Hugging Face. The model's weights are designed to effortlessly integrate with existing systems tailored for shorter contexts, particularly those that accommodate up to 2048 tokens. In addition to these features, we provide evaluation results and comparisons to the original OpenLLaMA models, thus offering a thorough insight into LongLLaMA's effectiveness in managing long-context tasks. This advancement marks a significant step forward in the field of language models, enabling more sophisticated applications and research opportunities.