The 7 best open-source LLMs to know in 2026

Takeaways

• The best open-source LLMs in 2026 are DeepSeek V4, GLM 5.2, Kimi K2.6, Kimi K3, MiniMax M3, Qwen3 VL 235B, and Google Gemma 4 31B.
• Kimi K3 posts the highest GPQA Diamond score of any open model at 93.5%, per the Onyx leaderboard, an independent benchmark aggregator, with full weights released July 27, 2026.
• The latest open source LLMs now match or beat proprietary models on reasoning and coding benchmarks. Mixture-of-experts architecture is the default across the field.
• DeepSeek V4 ships in two variants. Pro leads on coding at 80.6% SWE-Bench. Flash leads on browsing at 85.9% BrowseComp.
• Google Gemma 4 31B replaces Gemma 3 with 85.2% MMLU, 84.3% GPQA Diamond, and 80.0% SWE-Bench, all per the Onyx leaderboard.
• You can run leading open-source models through the Telnyx LLM Library with sub-500ms latency and no GPU management.

The best open-source LLMs in 2026

The best open-source LLMs in 2026 are DeepSeek V4, GLM 5.2, Kimi K2.6, Kimi K3, MiniMax M3, Qwen3 VL 235B, and Google Gemma 4 31B. These seven lead the latest open source LLMs on benchmarks and real-world use. Kimi K3, with full weights released July 27, posts 93.5% on GPQA Diamond. Gemma 4 31B brings dense-model performance to 85.2% MMLU and 80.0% SWE-Bench. DeepSeek V4 ships in two variants, Pro and Flash, both with 1M context windows. OpenAI also entered the open-source space with GPT oss 120b and 20b, though those models are not yet frontier-grade. The seven on this list are competitive with proprietary models on coding and math benchmarks, and you can self-host every one of them.

The field has moved fast. Two years ago, LLaMA 3 and Falcon topped most lists. Today, the frontier belongs to sparse mixture-of-experts models from labs in China, plus Google's efficient Gemma line. If you're new to the topic, start with what an LLM is before picking a model.

Comparing the best open-source LLMs

1. DeepSeek V4

DeepSeek V4 comes in two variants. Pro is the heavyweight. It scores 80.6% on SWE-Bench and 90.1% on GPQA Diamond. Flash trades a little raw power for speed and cost. It still holds its own with 79% on SWE-Bench and 88.1% on GPQA Diamond.

Flash has one trick Pro does not. It leads the field on BrowseComp at 85.9%. That benchmark tests how well a model finds hard-to-locate information on the web. If your workload involves agentic search, Flash is the pick.

Both variants use a Mixture-of-Experts architecture with 671B total parameters and 37B active per token. Both offer a 1M context window. Math remains a strength, with 96.0% on GSM8K. Licensing is permissive under Apache 2.0 and MIT, so commercial use is straightforward.

Best for: Teams that want frontier-level coding and reasoning without frontier-level costs. Pick Pro for depth. Pick Flash for agentic web tasks.

2. GLM 5.2

Zhipu AI's GLM 5.2 sits at the top of the Vellum leaderboard, and it earned the spot. The model packs 744B total parameters with 40B active per token, plus a 1M-token context window. That context size changes what's practical. You can feed it an entire codebase or a multi-hour transcript in a single request.

Its scores hold up across the hardest evaluations in use today:

• 54.7% on Humanity's Last Exam, the current frontier reasoning benchmark
• 91.2% on GPQA Diamond for graduate-level science questions
• 77.8% on SWE-Bench Verified (reported by Zhipu AI, not yet listed on Vellum)
• 1M-token context for long-horizon agentic tasks that span many steps

Best for complex systems engineering and agents that need to hold a lot of state over long sessions. Zhipu ships GLM under an MIT-style license. Check the model card on HuggingFace for the exact terms of your version.

3. Kimi K2.6

Moonshot AI's Kimi K2.6 is the largest model on this list at 1 trillion total parameters, though only 32B activate per token. It ranks second overall on Vellum and leads most computer-use evaluations. This is the model people reach for when they want an agent that can actually operate software in production.

The numbers back that up:

• 80.2% on SWE-Bench Verified, one of the best coding scores of any open model
• 73.1% on OSWorld, the top computer-use result in this group
• 90.5% on GPQA and 54% on Humanity's Last Exam
• 256K context window for long agent sessions

Best for autonomous agents, computer use, and coding. Moonshot releases K2 models under a modified MIT license, so read the terms before commercial deployment. In practice the restrictions are light.

4. Kimi K3

Moonshot AI released Kimi K3 in July 2026 and published full weights on July 27, 2026. The model uses a mixture-of-experts architecture with 2.8T total parameters, activating 16 of 896 experts per token, per Moonshot's model card. That keeps inference costs far below what the raw parameter count suggests.

The headline number is reasoning. Kimi K3 scores 93.5% on GPQA Diamond, the highest of any open model per the Onyx leaderboard. Its 1M-token context window handles entire codebases, long legal documents, or multi-session agent histories without truncation.

Moonshot publishes K3 under a modified MIT license with light commercial restrictions. Read the model card before deployment.

Best for: Reasoning-heavy workloads at frontier scale. If your application needs graduate-level scientific reasoning or long-document analysis with open weights, Kimi K3 is the current ceiling.

5. MiniMax M3

MiniMax M3 is the reasoning specialist. Its 93% on GPQA Diamond is the second-highest in this group, behind Kimi K3, and it pairs that with a 1M-token context window. It also posts 80.5% on SWE-Bench Verified, edging out Kimi K2.6, and 70.1% on OSWorld for computer use.

What stands out in daily use:

• 93% GPQA Diamond, the second-strongest reasoning score on this list
• 80.5% SWE-Bench Verified for agentic coding
• 1M-token context window matching GLM 5.2
• Strong performance on multi-step tasks where cheaper models lose the thread

Best for reasoning-heavy work and agentic coding pipelines. MiniMax publishes open weights under a permissive license. As with GLM, confirm the current terms on the model card before you ship.

6. Qwen3 VL 235B

Alibaba's Qwen3 VL 235B is the multimodal entry on this list. It's a 235B-parameter MoE model with 22B active per token, and it handles images and text in one model. Native context is 256K tokens, expandable to 1M. It scores 87.1% on MMLU and 90.3% on GSM8K per Alibaba's model card, so the vision capability doesn't come at the cost of language performance.

Where it earns its spot:

• Vision and language in a single model, no separate OCR pipeline needed
• Reads documents in 32 languages, including low-quality scans
• GUI automation support for agents that navigate screens
• Apache 2.0 license with full commercial rights

Best for visual comprehension, multilingual OCR, and GUI automation. If your product touches documents, screenshots, or interfaces, this is the open model to test first. It also pairs well with conversational AI applications that need to see what users see.

7. Google Gemma 4 31B

Google released Gemma 4 31B IT in April 2026 under an Apache 2.0 license, per Google's model card. Unlike the MoE giants on this list, it is a dense 32.2B-parameter model. That makes it practical to run on a single high-memory GPU or on-device hardware.

The jump from Gemma 3 is substantial. Context grew from 128K to 262K tokens. MMLU rose from 76.9% to 85.2%. GPQA Diamond sits at 84.3%. Gemma 4 posts an 80.0% SWE-Bench score, a benchmark Gemma 3 did not register on, per the Onyx leaderboard and Fireworks AI data.

Why small teams keep choosing it:

• Runs locally on one high-memory GPU, no cloud dependency
• 85.2% MMLU and 84.3% GPQA Diamond at 31B dense parameters
• Apache 2.0 license with full commercial rights
• 262K context window, double the predecessor's 128K

Best for local deployment and on-device workloads where data can't leave your infrastructure. The dense architecture and Apache 2.0 license make it the strongest small open model for regulated environments.

OpenAI enters open source

OpenAI entered the open-source space. The company released two open-weight models, GPT oss 120b and GPT oss 20b. The larger model scores 14.9% on Humanity's Last Exam with a 131K context window. The smaller one scores 10.9%.

Neither model competes with DeepSeek or GLM on hard reasoning benchmarks. That is not the point. These are smaller, efficient models from the biggest name in the industry, and they run on modest hardware.

The signal matters more than the scores. When OpenAI publishes open weights, the market has shifted. The next release could be a real contender.

Open-source vs. closed-source LLMs

Open-source models give you the weights. You can inspect them, fine-tune them, self-host them, and switch providers whenever you want. Closed models like GPT and Claude give you an API and nothing else. You trade control for convenience.

The performance gap has mostly closed. GLM 5.2 and Kimi K3 are competitive with proprietary frontier models on SWE-Bench and GPQA Diamond. The remaining differences come down to operations, not capability:

• Open models win on cost at scale, data privacy, and customization through fine-tuning
• Closed models win on zero-setup deployment and managed safety tooling
• Hybrid stacks are common. Teams route routine traffic to open models and reserve proprietary APIs for edge cases

For most production workloads in 2026, an open model behind a managed inference API is the sensible default.

How to choose the right open-source LLM

Match the model to the job, not the leaderboard. A 1T-parameter model is the wrong answer if your workload fits on a 31B model running locally.

Work through these questions in order:

1. What's the task? Coding agents point to Kimi K2.6 or MiniMax M3. Reasoning at frontier scale points to Kimi K3. Documents and images point to Qwen3 VL.
2. How much context do you need? Codebase-scale inputs need DeepSeek V4, GLM 5.2, or Kimi K3 and their 1M-token windows.
3. Where will it run? If data must stay on your hardware, Gemma 4 31B is the only model here that fits a single high-memory GPU.
4. What does the license allow? DeepSeek V4 and Qwen3 VL ship under Apache 2.0. Others use MIT variants. Read the terms before committing.
5. What's your latency budget? Voice applications need fast time-to-first-token. Smaller active parameter counts and co-located infrastructure matter more than raw benchmark scores here.

Cross-check candidates on the HuggingFace leaderboard with your own eval set. Public benchmarks tell you a model is competent. Only your data tells you it fits.

Future trends in open-source LLMs

The direction is clear. Open models are catching closed ones, and the gap closes every quarter.

Here is what to watch in 2026 and beyond:

• MoE has won. Nearly every top open model uses a Mixture-of-Experts architecture. Massive total parameters, small active counts. It is the only way to get frontier performance at open-source prices.
• Agentic benchmarks are replacing trivia tests. MMLU told you what a model knew. New benchmarks like BrowseComp and Terminal-Bench tell you what a model can do. Web research, terminal work, multi-step tasks. That is where the competition lives now.
• The market is converging toward open. OpenAI released GPT oss 120b and 20b. When the last major closed lab publishes open weights, the debate is over. Open is the default.
• Voice is becoming first-class. Models are shipping with native audio in and audio out. Text-only pipelines will look dated by 2027.

The takeaway is simple. Pick the model that fits your task, not the one with the biggest headline number. Test it on your own workload before you commit.

Open models are getting fast enough for real-time conversation. Platforms now pair them with Voice AI infrastructure and a Speech-to-Text API on the same network. If you want to see what that looks like in practice, here's how to build an AI voice agent end to end.

Frequently asked questions