Best Open Source LLMs in 2026

Two years ago, using an open-source LLM for serious work meant accepting a meaningful capability gap versus GPT-4 or Claude. That's no longer true. In 2026, MIT-licensed models like DeepSeek-V4-Pro and Kimi K2.6 now approach proprietary frontier models on several coding and reasoning benchmarks. For teams with data privacy requirements, the need to fine-tune on their own data, or the desire to avoid recurring API costs, the open-source tier is now a viable primary choice, not just a fallback.

This guide covers the top 10 open-source and open-weight LLMs from the Onyx Open LLM Leaderboard, updated as of July 20, 2026.

How this guide is sourced: Licensing, benchmark, parameter, and API availability data comes from the Onyx Open LLM Leaderboard. The recommendations in each section are editorial guidance for teams comparing open-source and open-weight options.

---

TL;DR: Open-source LLMs have closed most of the gap with proprietary models for coding and reasoning tasks. DeepSeek-V4-Pro and Kimi K2.6 are the two strongest picks: DeepSeek-V4-Pro leads on SWE-bench Verified and LiveCodeBench under an MIT license, while Kimi K2.6 posts a 90.5% GPQA Diamond, second only to GLM-5.2's 91.2% in this snapshot. For teams that need Apache 2.0 licensing, Qwen3.6-27B is a strong reasoning and coding option that runs on a single RTX 4090. If you want a cheap hosted API rather than self-hosting, DeepSeek-V4-Flash at $0.14/M input is the best reference point. For teams that want to run a capable model on modest hardware, Gemma 4 31B is the practical choice. Moonshot's Kimi K3 is already live via API with Arena Elo 1,486, but its weights aren't due until July 27, 2026, so it isn't ranked here yet.

---

What Is an Open-Source LLM?

An open-source large language model is one whose weights are publicly available for download, so you can run it on your own hardware, fine-tune it on your own data, and deploy it without paying per-token API fees. The license determines what you can actually do with it commercially.

The most permissive licenses are MIT and Apache 2.0, which allow unrestricted commercial use. The Llama License (Meta) and Gemma License (Google) are open for most uses but have specific restrictions.

License types in this guide:

---

What Is Onyx?

Onyx is an open-source AI platform that helps teams operationalize open-source and open-weight LLMs. The model provides inference, while Onyx supplies connectors, permission-aware retrieval, AI chat, agents, deep research, and user access across the organization.

That matters because adopting an open model is rarely the end of the project. Teams still need to connect it to Slack, Confluence, Google Drive, Jira, GitHub, and other knowledge sources without breaking access controls. Onyx provides that layer while keeping model choice open.

Best Open Source LLMs 2026: Comparison Table

Source: Onyx Open LLM Leaderboard, last updated July 20, 2026.

---

Top Open Source Models at a Glance

Best at fixing real bugs: DeepSeek-V4-Pro, MiniMax M3, Kimi K2.6, DeepSeek-V4-Flash, Hunyuan Hy3

Best code generation: DeepSeek-V4-Pro, DeepSeek-V4-Flash, Kimi K2.6, Nemotron 3 Ultra, Qwen3.6-27B

Best reasoning: GLM-5.2, Kimi K2.6, Hunyuan Hy3, DeepSeek-V4-Pro, DeepSeek-V4-Flash

Best on Humanity's Last Exam: GLM-5.2, Hunyuan Hy3, DeepSeek-V4-Flash, Nemotron 3 Ultra

Low-cost API options: Gemma 4 31B at $0/M (open weights), DeepSeek-V4-Flash at $0.14/M, Step-3.7-Flash at $0.20/M

---

Top Open Source Models: Detailed Reviews

1. DeepSeek-V4-Pro (DeepSeek, MIT)

Facts: DeepSeek-V4-Pro is listed as MIT-licensed with 1.6T total / 49B active parameters, 80.6% SWE-bench, 90.1% GPQA Diamond, 93.5% LiveCodeBench, 67.9% Terminal-Bench 2.0, and costs $0.435 / $0.87 per 1M input/output tokens.

Recommendation: Choose DeepSeek-V4-Pro when you want the strongest open-weight combination of software-engineering and code-generation performance under a fully permissive license.

2. Kimi K2.6 (Moonshot, Modified MIT)

Facts: Kimi K2.6 is listed under a Modified MIT license with 1T total / 32B active parameters, 80.2% SWE-bench, 90.5% GPQA Diamond, 89.6% LiveCodeBench, 83.2% BrowseComp, 73.1% OSWorld, 66.7% Terminal-Bench 2.0, and costs $0.95 / $4 per 1M input/output tokens.

Recommendation: Pick Kimi K2.6 if your top priority is near-frontier reasoning with the broadest agentic-benchmark coverage (BrowseComp, OSWorld, Terminal-Bench 2.0) in this comparison, and you can accept its modified MIT terms.

3. GLM-5.2 (Zhipu AI, MIT)

Facts: GLM-5.2 is listed as MIT-licensed with 753B total / 40B active parameters, 91.2% GPQA Diamond, 40.5% on Humanity's Last Exam, 1,468 Arena Elo, and costs $1.40 / $4.40 per 1M input/output tokens.

Recommendation: GLM-5.2 is the better fit than the coding-focused picks above when general reasoning and human-preference win rate matter more than a published SWE-bench score.

4. MiniMax M3 (MiniMax, MiniMax Community)

Facts: MiniMax M3 is listed under the MiniMax Community license with 428B total / 23B active parameters, 80.5% SWE-bench, 78.1% MMMU-Pro, 83.5% BrowseComp, 1,445 Arena Elo, and costs $0.30 / $1.20 per 1M input/output tokens.

Recommendation: MiniMax M3 is a cost-driven choice for teams comfortable reviewing its community license terms before production deployment.

5. DeepSeek-V4-Flash (DeepSeek, MIT)

Facts: DeepSeek-V4-Flash is listed as MIT-licensed with 284B total / 13B active parameters, 79.0% SWE-bench, 88.1% GPQA Diamond, 91.6% LiveCodeBench, 34.8% on Humanity's Last Exam, 56.9% Terminal-Bench 2.0, and costs $0.14 / $0.28 per 1M input/output tokens.

Recommendation: DeepSeek-V4-Flash is the strongest low-cost option here when you want most of DeepSeek-V4-Pro's capability at a fraction of the price.

6. Hunyuan Hy3 (Tencent, Apache 2.0)

Facts: Hunyuan Hy3 is listed as Apache 2.0 licensed with 295B total / 21B active parameters, 78.0% SWE-bench, 90.4% GPQA Diamond, 37.0% on Humanity's Last Exam, 84.2% BrowseComp, and 1,412 Arena Elo. Pricing is not listed in this snapshot.

Recommendation: Choose Hunyuan Hy3 when you want Apache licensing with strong reasoning and agentic scores and plan to self-host rather than rely on a published API price.

7. Step-3.7-Flash (Stepfun, Apache 2.0)

Facts: Step-3.7-Flash is listed as Apache 2.0 licensed with 198B total / 11B active parameters, 76.5% SWE-bench, 75.8% BrowseComp, and costs $0.20 / $1.15 per 1M input/output tokens.

Recommendation: Step-3.7-Flash is useful as a budget-friendly, Apache-licensed pick when you want solid SWE-bench performance without the largest parameter counts in this list.

8. Qwen3.6-27B (Alibaba, Apache 2.0)

Facts: Qwen3.6-27B is listed as Apache 2.0 licensed with 27B total / 27B active parameters, 77.2% SWE-bench, 87.8% GPQA Diamond, 83.9% LiveCodeBench, 59.3% Terminal-Bench 2.0, and costs $0.60 / $3.60 per 1M input/output tokens.

Recommendation: Qwen3.6-27B is a strong choice when you want Apache licensing with dense-model reliability and a footprint small enough to run on a single RTX 4090.

9. Nemotron 3 Ultra (Nvidia, OpenMDW-1.1)

Facts: Nemotron 3 Ultra is listed under Nvidia's OpenMDW-1.1 license with 550B total / 55B active parameters, 70.7% SWE-bench, 87.0% GPQA Diamond, 89.0% LiveCodeBench, 81.7% IFEval, 26.7% on Humanity's Last Exam, and 70.9% Tau2-Bench. Pricing is not listed in this snapshot.

Recommendation: Nemotron 3 Ultra is a reasonable pick when you want an Nvidia-optimized open model with broad benchmark coverage and plan to self-host.

10. Gemma 4 31B (Google, Gemma)

Facts: Gemma 4 31B is listed under the Gemma license with 31B total / 31B active parameters, 84.3% GPQA Diamond, 80.0% LiveCodeBench, 88.4% MMMLU, 1,451 Arena Elo, and costs $0 / $0 per 1M input/output tokens.

Recommendation: Gemma 4 31B is the strongest option in this list when you want a free, single-GPU-friendly model rather than reaching for the largest parameter counts here.

---

Open Source LLMs: Best for Each Use Case

---

Using Open-Source LLMs in Enterprise Workflows

Choosing an MIT or Apache 2.0 model is only part of the decision. Teams also need a way to compare hosted and self-hosted backends, connect those models to internal data, and preserve permissions.

Onyx is useful in this context because it gives teams a common application layer on top of open-source models. You can test a hosted API against a self-hosted endpoint, connect the chosen model to sources like Slack, Confluence, Jira, Google Drive, and GitHub, and keep permission-aware retrieval in front of users. That makes it easier to act on the licensing and deployment tradeoffs in this guide instead of evaluating each model in isolation.

---

Recommended Open-Source LLM Stack

Pick the model and platform together. A permissive model license is valuable only if the surrounding application layer also supports the deployment and governance requirements.

Frequently Asked Questions

What is the best open-source LLM in 2026?

In this leaderboard snapshot, there is no single winner across every benchmark. DeepSeek-V4-Pro leads SWE-bench (80.6%) and LiveCodeBench (93.5%). GLM-5.2 has the highest GPQA Diamond score among the open-weight models listed (91.2%). Kimi K2.6 is close behind on GPQA Diamond (90.5%) while also posting strong BrowseComp and OSWorld agentic scores. The best choice depends on whether you care most about coding, reasoning, licensing, or deployment constraints. Moonshot's Kimi K3 is worth watching too: it's already available via API with an Arena Elo of 1,486, but its weights aren't scheduled to ship until July 27, 2026.

What is the difference between open-source and open-weight LLMs?

Open-source LLMs have publicly available weights, architecture, and (ideally) training code under a permissive license like MIT or Apache 2.0. Open-weight models release weights but may have proprietary training code or restrictive license terms. In practice, most "open-source" LLMs are open-weight: you can download and run them, but full source code and training data are rarely published.

Which open-source LLMs can I use commercially?

MIT and Apache 2.0 licensed models allow commercial use without restriction: DeepSeek-V4-Pro, DeepSeek-V4-Flash, GLM-5.2, Hunyuan Hy3, Step-3.7-Flash, Qwen3.6-27B. Kimi K2.6 ships under a Modified MIT license and Nemotron 3 Ultra under Nvidia's OpenMDW-1.1, both of which require reading the specific terms before commercial deployment. MiniMax M3 uses the MiniMax Community license, which has its own commercial-use conditions. Gemma 4 31B uses the Gemma license, which permits commercial use but prohibits uses that harm Google products. Always check the specific license terms for your deployment scenario.

Can I run these open-source models locally?

Most models in this list require enterprise hardware (4x H100 80GB or more) for full-precision inference, and the largest, like DeepSeek-V4-Pro, GLM-5.2, and Kimi K2.6, need 16x H100 80GB. More accessible options include Qwen3.6-27B and Gemma 4 31B, which both run on a single RTX 4090 24GB. See the Best Self-Hosted LLMs 2026 guide for hardware requirements per model.

What is the best platform for running open-source and proprietary LLMs together?

Most teams end up mixing models: a self-hosted open-weight model for sensitive data, a cheap API for high-volume tasks, and a frontier model for the hardest work. Onyx gives teams a single interface to connect all of these, routing tasks to the right model while keeping answers grounded in company knowledge from Slack, Confluence, Jira, Google Drive, and GitHub. It's MIT-licensed, supports self-hosted and API-based backends, and is free to get started.