On
Comparing 120+ Open-Source LLMs: The Best Models in 2026

Open-source large language models have become one of the most significant AI trends heading into 2026. And honestly, that makes sense: For years, open-weight models lagged far behind proprietary alternatives. But that gap has shrunk dramatically.

Recent releases show the shift clearly. DeepSeek V4 Pro (released April 24, 2026), GLM-5.1 from Z.ai, Kimi K2.6 from Moonshot AI, and Qwen3.5 from Alibaba now rival the best proprietary LLMs—including Claude Opus 4.8, GPT-5.5, and Gemini 3.1 Pro. On specific benchmarks like SWE-Bench Pro and HumanEval, open models actually outperform them.

The new proprietary baseline for Terminal-Bench 2.1 is GPT-5.6 Sol (widely released July 9, 2026; OpenAI's default model in Codex) at 88.8%, with Sol Ultra reaching 91.9%. Sol scores 64.6% on SWE-Bench Pro—about 15 points below Claude Mythos 5 and Fable 5. Other open-source equivalents on Terminal-Bench 2.1 haven't been published yet.

This guide contains a curated list of 120+ open-source LLMs, complete with benchmarks, licenses, API pricing, context windows, and capabilities.

Complete Overview of All Open-Source LLMs

The table below includes every open-weight model from models.dev plus hand-picked classics, sorted by release date (newest first). It displays additional metrics like method, knowledge cutoff, max output, and number of API providers:

Model Developer Parameters Context MMLU Math Code Reasoning Tools License Input Price Output Price Release Date
Laguna XS 2.1 Poolside 33B (3B active) 256K – – 70.9% SWE-Bench Verified OpenMDW-1.1 $0.060 $0.12 July 2026
DeepSeek V4 Pro DeepSeek 1.6T (49B active) 1M 87.5% MMLU-Pro 90.1% GPQA 93.5% LiveCodeBench MIT $0.35 $0.70 April 2026
Kimi K2.6 Moonshot AI 1T (32B active) 256K 84.6% MMLU-Pro 90.5% GPQA 92% HumanEval Modified MIT $0.28 $1.10 April 2026
GLM-5.1 Z.ai 754B 200K 91.7% MMLU 85.7% GPQA 58.4% SWE-Bench Pro MIT $0.30 $2.15 April 2026
Llama 3.1 405B Instruct Meta 405B 128K 88.6% MMLU 73.8% MATH 89% HumanEval Llama 3.1 Community – – July 2024
DeepSeek-V3 DeepSeek 671B (37B active) 128K 88.5% MMLU 90.2% MATH 85% HumanEval MIT $0.25 $1.00 December 2024

The Best Open-Source Models in 2026

Let's look ahead first: On July 16, 2026, Moonshot AI launched Kimi K3, a mixture-of-experts model with 2.8 trillion parameters (896 experts, 16 active per token) and a 1-million-token context window. It supports native image input and scores 93.5% on GPQA Diamond. Currently K3 is only available via app and API ($3/$15 per million tokens; also on OpenRouter). Weights arrive on Hugging Face on July 27, 2026 (Moonshot will announce the exact license at release; previous models use modified MIT). This will make K3 the largest open-weight model ever released. Once weights drop, it'll slot into the table below. Until then, the April rankings apply:

DeepSeek V4 Pro (April 24, 2026) is the new leader. This mixture-of-experts model with 1.6 trillion parameters activates just 49 billion tokens. It scores 87.5% on MMLU-Pro, 90.1% on GPQA Diamond, and 93.5% on LiveCodeBench. It uses the same MIT license as other DeepSeek products and ships with native 1-million-token context plus reasoning performance at about 27% of the v3.2 version.

Kimi K2.6 from Moonshot AI ranks second for open performance: 92% on HumanEval, 90.5% on GPQA Diamond, 96.4% on AIME 2026, with 256K context and native video input. Mixture-of-experts with 1 trillion parameters under a modified MIT license.

GLM-5.1 from Z.ai (formerly Zhipu) leads SWE-Bench Pro at 58.4%—beating GPT-5.4 (57.7%) and Claude Opus 4.6 (57.3%). This mixture-of-experts with 754 billion parameters was trained entirely on Huawei Ascend chips and ships under MIT license. Its reasoning-focused sibling, GLM-5, hits 96% on MMLU and 94% on GPQA—the highest knowledge scores in open source.

Kimi K2.5 still owns the highest HumanEval score across all benchmarks (99.0) and leads MATH-500 (98.0). This is the best open-source model for code generation when latency matters less than quality.

DeepSeek V4 Flash (284B / 13B active) is the cost-efficient V4 Pro variant and the smartest pick when you want top-tier performance on a single high-end GPU.

Previous generations remain capable: GPT-OSS-120B (OpenAI's first open weights since GPT-2), DeepSeek R1, Qwen3-235B-A22B-Thinking, and Llama 4 Maverick are all still strong—just no longer cutting-edge.

Here's a side-by-side of the top 5:

Feature DeepSeek V4 Pro Kimi K2.6 GLM-5.1 Kimi K2.5 DeepSeek V4 Flash
Developer DeepSeek Moonshot AI Z.ai Moonshot AI DeepSeek
License MIT Modified MIT MIT Modified MIT MIT
Total Parameters (MoE) 1.6T (49B active) 1T 754B 1T 284B (13B active)
Run Locally Yes Yes Yes Yes Yes
Local Hardware Multiple GPUs or quantization Multiple GPUs or quantization Multiple GPUs or quantization Multiple GPUs or quantization Single high-end GPU
Standout Feature Native 1M-token context 256K context, native video Trained on Huawei Ascend MATH-500 leader at 98.0 Cost-efficient V4 Pro variant

Description: Comprehensive guide to 120+ open-source large language models with benchmarks, pricing, licenses, and context windows for 2026.

Related Articles