Modelos de programación con pesos abiertos

Inkling

Thinking Machines Lab's open-weights general-purpose multimodal Mixture-of-Experts model with 975B total parameters and 41B active parameters. Inkling accepts text, image, and audio inputs, produces text, and is designed for agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation.

Kimi K3

Moonshot AI's flagship Kimi model for frontier intelligence, agentic coding, knowledge work, and deep reasoning. Kimi K3 supports a 1-million-token context window for long-running software engineering and research workflows.

Sarvam-105B

Sarvam AI's sovereign 105B-parameter Mixture-of-Experts model activating ~9B parameters per token, with a 128K-token context window. Trained on 12 trillion tokens across 22 Indian languages using 128 sparse experts with Multi-head Latent Attention and a custom low-fertility Indic tokenizer. Wins the majority of pairwise comparisons on Indian-language and STEM benchmarks.

Sarvam-30B

Sarvam AI's 30B-parameter Mixture-of-Experts reasoning model trained from scratch with only 2.4B active parameters per token. Optimized for real-time deployment and Indian languages, delivering strong reasoning, coding, and conversational performance while remaining efficient to serve. Open-weights.

Sarvam-M

Sarvam AI's 24B-parameter instruction-tuned model derived from Mistral-Small-3.1-24B, post-trained on English plus eleven major Indic languages (bn, hi, kn, gu, mr, ml, or, pa, ta, te). Delivers large relative gains on Indian-language, math, and programming benchmarks over its base model, with a hybrid reasoning mode for complex tasks.

Command A+

Cohere's enterprise flagship model building on Command A with stronger reasoning, agentic tool use, and multilingual performance across 23 languages. Optimized for secure, high-throughput RAG, retrieval, and long-horizon agent workflows in regulated environments, with private and on-premise deployment options.

GLM-5.2

Z.ai's (formerly Zhipu AI) flagship open-weight coding model with a 1M-token context window. Mixture-of-Experts architecture with 753B total parameters and ~40B active per request, featuring two cost-balancing reasoning modes. Tops several coding benchmarks while remaining a fraction of the cost of comparable proprietary frontier models. MIT-licensed weights.

Kimi K2.7 Code

Moonshot AI's latest open-source, coding-focused model in the Kimi K2 family, built to complete end-to-end programming tasks reliably over long contexts. A 1-trillion-parameter model that cuts reasoning token usage by roughly 30% versus K2.6 while improving coding and agent performance — +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite for multi-language support. Released under a Modified MIT License and available via Kimi APIs and Hugging Face.

DiffusionGemma

Google DeepMind's experimental diffusion-based member of the Gemma 4 open model family. Unlike autoregressive models that generate text one token at a time, DiffusionGemma denoises a canvas of placeholder tokens to produce up to 256 tokens in parallel, finalizing output in one block. A Mixture-of-Experts model with 26B total parameters and 3.8B active per inference, delivering roughly 4x the throughput of similarly sized autoregressive Gemma models on local hardware. Excels at non-linear tasks like in-line editing, molecular sequencing, mathematical graphing, and self-correcting puzzles.

Nemotron 3 Ultra

NVIDIA's flagship open 550B-parameter Mixture-of-Experts model with 55B active parameters, built for frontier reasoning and orchestration in long-running agentic systems. Features hybrid Mamba-Transformer architecture, LatentMoE routing, multi-token prediction, and NVFP4 precision for 5x higher throughput. Achieves 30% lower cost-to-task-completion on agentic benchmarks. Supports 1M+ token context window with 95% accuracy on Ruler@1M.

Gemma 4 12B

Google's medium-size open-weight model with 12 billion parameters from the Gemma 4 family. Encoder-free unified multimodal architecture that natively processes text, image, audio, and video inputs without dedicated encoders. Features a 256K context window and supports 140+ languages. First medium-sized model capable of natively ingesting audio. Suitable for local deployment on GPUs with 16GB VRAM.

DBRX

Databricks' open-source 132B parameter Mixture-of-Experts transformer model with 36B active parameters per input. Released under Databricks Open Model License, optimized for enterprise workloads including SQL generation and coding tasks.

Jamba Large 1.7

AI21's latest hybrid SSM-Transformer model with Mixture-of-Experts architecture. Features a 256K context window, improved grounding and instruction-following. 94B total parameters with 398B active, optimized for enterprise long-context tasks.

Falcon 3 10B

TII's open-source 10B parameter model from the Falcon 3 family. Achieved number one position on Hugging Face's LLM leaderboard in its size category, outperforming Meta's Llama variants and other models under 13B parameters.

Snowflake Arctic

Snowflake's enterprise-focused open LLM with 480B total parameters using a fine-grained MoE architecture with only 17B active parameters per input. Apache 2.0 licensed, excels at SQL generation, coding, and enterprise intelligence tasks with breakthrough training efficiency.

StableLM 2 12B

Stability AI's 12.1 billion parameter decoder-only language model pre-trained on 2 trillion tokens of diverse multilingual and code datasets. Supports multiple languages and offers strong performance for its compact size with instruction-tuned chat variant available.

Falcon-H1

TII's hybrid Mamba-Transformer model that outperforms comparable offerings from Meta's Llama and Alibaba's Qwen in the 30-70B parameter range. Designed for real-world AI on everyday devices and resource-limited settings with state-of-the-art efficiency.

Ring-2.6-1T

InclusionAI's (Ant Group) trillion-parameter open-weights reasoning model with 63B active parameters per token. Built for real-world agent workflows with adaptive reasoning-effort modes. Features hybrid linear and MLA attention architecture with MIT license.

K2 Think

A 32 billion parameter open-weights reasoning model by LLM360/MBZUAI, built on Qwen2.5-32B. Trained with reinforcement learning and verifiable rewards for long chain-of-thought reasoning, agentic planning, and complex problem solving in math, science, and code.

Granite 4.1 8B

IBM's dense decoder-only 8B parameter language model from the Granite 4.1 family. Supports 131K-token context, tool calling, RAG, code generation with fill-in-the-middle, text summarization, classification, and extraction across 12 languages. Released under Apache 2.0.

Granite 4.1 30B

IBM's largest dense decoder-only 30B parameter language model from the Granite 4.1 family. Trained on approximately 15T tokens with long-context extension up to 512K tokens. Supports tool calling, RAG, code generation, multilingual tasks across 12 languages. Released under Apache 2.0.

DeepSeek V4 Pro

DeepSeek's flagship V4 model with 1.6T total parameters (49B activated). MoE architecture supporting 1M token context. Closes the gap with frontier proprietary models on reasoning and coding benchmarks.

DeepSeek V4 Flash

DeepSeek's efficient V4 model with 284B total parameters (13B activated). Optimized for speed and cost-efficiency while maintaining strong performance. Supports 1M token context window.

Qwen 3.6 27B

Alibaba's dense 27B parameter model that outperforms its own 397B MoE predecessor on agentic coding benchmarks. Strong multilingual and reasoning capabilities released under Apache 2.0.

MiMo-V2.5-Pro

Xiaomi's flagship 1.02T-parameter Mixture-of-Experts model with 42B active parameters, built on a hybrid-attention architecture with 3-layer Multi-Token Prediction. Designed for complex agentic tasks, software engineering, and long-horizon instruction following with a 1M-token context window.

Hy3 Preview

Tencent's flagship open-weight Mixture-of-Experts model from the Hunyuan family with 295B total parameters and 21B active. Integrates fast and slow thinking modes with configurable reasoning effort. Designed for agentic workflows, cross-file code refactoring, long-document analysis, and multi-step tool use.

Qwen 3.6 35B-A3B

Alibaba's efficient Mixture-of-Experts model with 35B total parameters and 3B active per token. Frontier-level agentic coding performance with 73.4% on SWE-bench Verified and 92.7 on AIME 2026. Released under Apache 2.0.

Gemma 4 E4B

Google's efficient 4 billion parameter variant from the Gemma 4 family. Designed for resource-constrained environments while maintaining strong text generation quality. Text-only model with a 32K context window, balancing performance and efficiency.

Gemma 4 26B

Google's high-performance open-weight dense model with 26 billion parameters from the Gemma 4 family. Supports multimodal inputs including text and images with a 256K extended context window. Strong reasoning and code generation capabilities with all parameters active per forward pass.

Gemma 4 31B

Google's flagship open-weight dense model with 31 billion parameters from the Gemma 4 family. All parameters active per forward pass with top-tier performance on reasoning benchmarks including AIME 2026 and MMLU Pro. Supports vision and extended 256K context window.

Gemma 4 31B

Google's flagship open-weight dense model with 31B parameters. All parameters active per forward pass. Ranks among top open models with strong performance on AIME 2026 (89.2%) and MMLU Pro (85.2%). Supports vision and extended context.

GPT-OSS 20B

OpenAI's compact open-weight model with 20 billion parameters. Released under Apache 2.0 license, designed for efficient deployment on consumer hardware while maintaining strong coding and reasoning capabilities.

Nemotron 3 Super 120B

NVIDIA's open hybrid Mamba-Transformer MoE model with 120B total parameters (12B active). Features 1M token context window and excels at agentic reasoning, coding, planning, and tool calling.

GPT-OSS 120B

OpenAI's first open-weight large model with 120 billion parameters. Released under Apache 2.0 license, offering strong performance on reasoning and coding tasks while being fully self-hostable.

Qwen 3.6

Alibaba's latest Qwen model with enhanced reasoning, multilingual capabilities, and improved instruction following. Features strong performance on coding, math, and general knowledge benchmarks.

Mistral Small 4

Mistral AI's efficient hybrid model unifying instruct, reasoning, and coding in a single model. Open-weight under Apache 2.0 with strong performance for its size class.

DeepSeek V4

DeepSeek's fourth-generation model with improved mixture-of-experts architecture, enhanced reasoning and coding capabilities, and stronger multilingual performance. Competitive with frontier proprietary models.

GigaChat 3.1 Lightning

Sber's compact Mixture-of-Experts model with 10B total parameters and 1.8B active. Designed for fast multilingual assistant workloads, reasoning, code, function calling, and product-style deployment on edge devices.

Devstral 2

Mistral AI's frontier code agents model designed for solving software engineering tasks. Open-weight model optimized for agentic coding workflows and complex development tasks.

Mistral Large 3

Mistral AI's largest open-weight model with 41B active parameters (675B total MoE). State-of-the-art general-purpose multimodal model with 256K context window and powerful agentic capabilities. Released under Apache 2.0.

GigaChat 3.1 Ultra

Sber's flagship large-scale Mixture-of-Experts model with 702B total parameters and 36B active. Designed for multilingual assistant workloads, reasoning, code generation, tool use, and large-cluster deployment. Open-weight release.

GLM-4.7

Zhipu AI's multilingual agentic coding model with strong reasoning, tool use, and UI generation capabilities. Predecessor to GLM-5.1 with competitive performance on coding benchmarks.

AlemLLM

Kazakhstan's flagship Mixture-of-Experts language model developed by Astana Hub with technical support from 01.AI. Features 247B total parameters with 22B active per token, achieving state-of-the-art results on Kazakh, Russian, and English benchmarks. Outperforms GPT-4o on Kazakh language tasks.

Trendyol LLM 8B T1

Turkish-optimized 8B chat model developed by Trendyol, Turkey's largest e-commerce platform. Built on Qwen3-8B and fine-tuned on large-scale Turkish e-commerce datasets. Features advanced chain-of-thought reasoning in Turkish with dual operation modes (/think and /no_think), strong instruction following, summarization, coding, and attribute extraction for catalogue enrichment. English reasoning capabilities are preserved alongside Turkish.

YandexGPT 5 Lite

Yandex's compact 8B parameter language model trained on 15T tokens of primarily Russian and English text. Features 32K context window with strong performance on web, code, and mathematics tasks. Open-weight release.

Nemotron Nano 9B v2

NVIDIA's compact 9B parameter model trained from scratch for both reasoning and non-reasoning tasks. Generates reasoning traces before final responses. Efficient for edge and on-device deployment.

Kumru 7B

Turkish large language model developed by VNGRS, pre-trained from scratch on 500 GB of Turkish corpora (300B tokens). Decoder-only architecture with a custom tokenizer optimized for Turkish, supporting code, math, and chat. Outperforms significantly larger multilingual models on the Cetvel Turkish benchmark. Available for on-premise enterprise deployment.

Llama 4 Maverick

Meta's quality-focused MoE model with 17B active parameters (400B total, 128 experts). Targets quality-critical tasks with benchmark scores competitive with GPT-4o and Gemini 2.5 Pro.

Llama 4 Scout

Meta's efficient MoE model with 17B active parameters (109B total, 16 experts). Supports up to 10M token context — the longest of any production model. Strong performance on reasoning and multilingual tasks.

Qwen3 32B

Alibaba's Qwen3 32B dense language model with strong reasoning and multilingual capabilities, supporting function calling and code generation across diverse tasks.

Qwen3 235B

Alibaba's Qwen3 235B mixture-of-experts model delivering frontier-level performance with advanced reasoning, function calling, and code generation capabilities at massive scale.

Qwen3 Coder

Alibaba's Qwen3 Coder model optimized for software development tasks including code generation, debugging, code review, and technical documentation with strong multilingual programming support.

Gemma 3 27B

Google's largest open-weight model in the Gemma 3 family with 27 billion parameters. Supports multimodal inputs including text and images with a 128K context window. Delivers strong performance across reasoning, code generation, and vision tasks, competitive with larger proprietary models.

Gemma 3 12B

Google's mid-size open-weight model with 12 billion parameters from the Gemma 3 family. Supports multimodal inputs including text and images with a 128K context window. Strong performance on reasoning and code generation tasks at moderate compute cost.

Gemma 3 4B

Google's compact open-weight model with 4 billion parameters from the Gemma 3 family. Supports multimodal inputs including text and images with a 128K context window. Balances efficiency and capability for vision and language tasks.

Command A

Cohere's flagship 111B parameter model optimized for demanding enterprises requiring fast, secure, and high-quality AI. Excels at RAG, tool use, and multilingual tasks with strong reasoning capabilities.

Mistral Small 3.1

Mistral AI's Small 3.1 model with 24B parameters offering efficient multimodal capabilities including vision, function calling, and code generation with a large 128K context window.

Phi-4 Mini

Microsoft's Phi-4 Mini model with 3.8B parameters providing lightweight yet capable language understanding and code generation, optimized for resource-constrained deployments with a large 128K context window.

DeepSeek R1

DeepSeek's reasoning-focused model trained with reinforcement learning for complex multi-step reasoning. Excels at math, science, and coding problems requiring chain-of-thought reasoning.

Codestral

Mistral AI's cutting-edge code generation model specializing in low-latency, high-frequency tasks such as fill-in-the-middle (FIM), code completion, correction, and test generation. Features efficient architecture with 2x faster generation than its predecessor.

ISSAI KazLLM 1.0 70B

Large language model developed by ISSAI (Nazarbayev University) customized from Llama 3.1 70B to improve helpfulness of responses in the Kazakh language. Part of Kazakhstan's initiative to ensure the country benefits from generative AI advancements.

Llama 3.3 70B Instruct

Meta's flagship open-weight model with 70 billion parameters. Strong multilingual capabilities with competitive performance on reasoning and coding benchmarks. Available for self-hosting and through various inference providers.

DeepSeek V3

DeepSeek's third-generation large language model featuring mixture-of-experts architecture, strong multilingual capabilities, and competitive performance on reasoning and coding benchmarks.

Phi-4

Microsoft's Phi-4 model with 14B parameters excelling at reasoning and code generation tasks, delivering strong performance relative to its compact size with efficient inference characteristics.

Command R7B

Cohere's compact 7B parameter model optimized for RAG, tool use, and code tasks. Delivers top-tier speed and efficiency on commodity GPUs and edge devices with 128K context window.

Llama 3.2 3B Instruct

Meta's lightweight open-weight model with 3 billion parameters from the Llama 3.2 family. Designed for on-device and edge deployment with strong text generation capabilities relative to its size. Supports instruction following and general-purpose tasks.

Llama 3.2 11B Vision Instruct

Meta's multimodal open-weight model with 11 billion parameters from the Llama 3.2 family. Supports both text and image inputs, enabling visual understanding tasks alongside standard text generation. Suitable for applications requiring vision capabilities at moderate scale.

Llama 3.2 90B Vision Instruct

Meta's largest multimodal open-weight model with 90 billion parameters from the Llama 3.2 family. Delivers strong performance on both text and image understanding tasks with competitive results on visual reasoning benchmarks. Designed for high-quality inference requiring vision capabilities.

Llama 3.1 8B Instruct

Meta's efficient open-weight model with 8 billion parameters from the Llama 3.1 family. Optimized for instruction following with strong performance on general tasks, coding, and multilingual benchmarks. Ideal for cost-effective deployment and edge inference scenarios.