Best Open-Source LLMs of 2026: Ranked by Capability, License, and Hardware

Last updated September 5, 2026

The best open-source LLMs in September 2026 are Qwen3.8-27B for anyone with a single 24 GB GPU or 32 GB Mac, GLM-5.3 and Qwen3.8-2.4T at the frontier-class top of the open-weight field, and Gemma 4 for machines under 16 GB. That ranking would have looked different a month ago: between August 10 and 29, six open-weight releases landed, two of them small enough for a consumer card and two of them large enough to need a cluster. This guide ranks the open-source AI models worth running now, explains what each license actually lets you do, and charts the strongest open weights against the closed frontier so you know precisely what you trade for owning the stack.

Key Takeaways

• Qwen3.8-27B is the best open-source LLM most people can run: Apache 2.0, multimodal, and a 61.7 percent SWE-bench Pro score on one consumer GPU, within a point of models that need eight data-center cards.
• The open-versus-closed gap on real-world coding is now single digits below Anthropic's top models. The open-weight Qwen3.8-2.4T scores 67.7 percent on SWE-bench Pro, ahead of GPT-5.6 and Claude Sonnet 5 on vendor numbers.
• "Open source" now means five different things. Apache 2.0 and MIT weights are clean; the Qwen3.8-Max, Kimi K3, LFM, and Llama 4 licenses each carry a revenue, user-count, or territory clause that matters the day you ship something commercial.
• Open weights are the hedge against policy. Claude Fable 5 went dark for 19 days in June under a US export-control order. Models people ran on their own hardware never noticed.

Open Source vs Open Weights: What You Are Actually Getting

Almost nothing on this page is open source in the strict sense. Open weights means the model's parameters are downloadable, so you can run it, fine-tune it, and keep your data off someone else's server. Open source, properly, also releases the training code and data under a license that lets you rebuild the thing from scratch. Only NVIDIA's Nemotron line publishes training data and post-training recipes alongside the weights; everyone else ships the finished model and keeps the kitchen closed.

For most purposes the distinction that matters more is the license on the weights. A model can be fully downloadable and still forbid commercial use above a revenue line, exclude a region, or require a separate agreement past a user count. The ranking below marks each model's terms, and the licensing section further down explains what each clause does.

The Best Open-Source LLMs Right Now, Ranked

The table sorts by the hardware each model needs, because that is the first filter for everyone, then by capability within each class. "One GPU" means a 24 GB card or a 32 GB Mac at 4-bit quantization. "128 GB" means a unified-memory workstation such as a Mac Studio or a Ryzen AI Max box. "Cluster" means multi-GPU server hardware that no household owns.

Benchmark figures are vendor-reported unless stated. Every model here has downloadable weights; the hardware column is what it takes to run them at 4-bit, not to download them.

Three things stand out in that table. First, the single-GPU class got three serious entrants in five days, and Qwen3.8-27B reset what the class can do. Second, MIT and Apache 2.0 now cover the entire open field from a phone to a cluster, which was not true in the spring. Third, the two largest models on the list, Qwen3.8-2.4T and Kimi K3, are the two with the most restrictive terms, so the very top of the open-weight leaderboard is also where the licensing gets careful.

What Fits on Your Hardware

The models above sort into four memory bands. This is the short version; the full tier-by-tier picks, quantized file sizes, and hardware recommendations are in our local AI models by VRAM guide, which is re-verified on a dated schedule.

The dense models (Qwen3.8-27B, Muse Glimmer) are bandwidth-bound on a Mac and feel slower than the mixture-of-experts options at the same size; if your workload is many short interactive steps rather than a few long ones, Nemotron 3.5 Lightning's 3B active parameters are a step change. Our Qwen3.8-27B hardware guide, DeepSeek V4-Flash reality check, and GLM-5.3-Flash reality check cover the specific machines, including the prefill speeds the demos skip.

How the Best Open Weights Compare to the Closed Frontier

The table charts the strongest open models against the closed frontier on SWE-bench Pro, the least saturated coding benchmark in wide use (real, recent GitHub issues). Claude Fable 5.1 replaced Fable 5 at the top of the vendor aggregate in early September.

Closed-frontier figures are from vendor system cards and the llm-stats aggregate. Open-model scores marked "vendor" are self-reported and not yet independently replicated. Z.ai has not published a separate SWE-bench Pro run for GLM-5.3, so its predecessor stays in the table. Read the ordering, not the decimals.

The row that matters is near the bottom. Qwen3.8-27B, a model you can run on a used RTX 3090, scores within a point of GLM-5.2 and above GPT-5.5 on vendor numbers. That is the first time a single-GPU open model has landed inside the frontier pack on this benchmark. The reasoning gap has narrowed too: Qwen reports 92.6 percent on GPQA Diamond for the 2.4T model against 93 to 94 for the closed frontier, where in the spring the best open figure was around 80. Treat every one of those as a claim until third parties replicate it. The takeaway is not that open weights have caught the frontier outright; Anthropic's top three still lead by more than ten points on the hardest coding and by a clear margin on long-horizon work. It is that for the agentic coding most teams run daily, the deciding factor is now hardware and licensing, not capability.

What the Licenses Actually Let You Do

This is where "open source AI" stops being one category. Five license patterns cover everything on this page, and the differences only bite when you ship something.

Apache 2.0 and MIT are the clean class: use, modify, redistribute, and sell, with attribution and no other conditions. Gemma 4, Qwen3.8-27B, Muse Glimmer, DeepSeek V4, GLM-5.2, GLM-5.3-Flash, and GLM-5.3 all sit here. If you cannot afford a legal review, stay in this column.

OpenMDW-1.1, used by NVIDIA's Nemotron 3 Ultra and 3.5 Lightning, is a permissive license that also covers the training data and recipes NVIDIA publishes. It is the closest thing on this list to open source in the full sense, and it is the right choice when a compliance team needs to see how a model was made.

Revenue-gated licenses are permissive until you are big. The Qwen3.8-Max License on the 2.4T weights requires AI-service businesses above $50 million in annual revenue to negotiate separately; Liquid AI's LFM license on its small models sets the line at $10 million; Kimi K3's custom license is gated the same way. For personal use, homelabs, and most small businesses these change nothing. Note that the two Qwen3.8 models arrived in one announcement under two different licenses, so check the LICENSE file in the repository, not the badge on the card.

User-count and territory clauses are the Llama 4 pattern: a separate license above 700 million monthly active users and an exclusion for EU-domiciled users. The first clause is irrelevant to nearly everyone; the second is a hard stop for a European deployment.

Custom community licenses, such as MiniMax's, need to be read in full. Its predecessor restricted commercial use without written authorization. A model under one of these can be excellent and still be the wrong choice for a product.

Privacy and Security: Local Is Not Automatic

Running a model on your own hardware is the necessary first step for digital sovereignty, not the sufficient one. A local stack that binds to 0.0.0.0 on your main LAN with no isolation is not meaningfully more private than a cloud service, because anything on your network can reach it.

The baseline: keep Ollama on its default localhost binding, run inference in a Docker container, put AI workloads on their own VLAN if your router supports it, monitor outbound DNS with Pi-hole, and pull weights only from verified repositories. For a larger build, see our guide to a private knowledge base with Obsidian and Ollama. Most of that VLAN and firewall work happens on your router, and if you are still using the gateway your ISP rents you, you may not have the settings to do it. Owning your equipment is the same principle as owning your model.

The Bigger Picture: When the Frontier Gets Switched Off

For most of 2026 the case for open-source AI was privacy, cost, and offline reliability. In June it gained a sharper edge, and the episode is now complete enough to judge. On June 12, 2026, a US Commerce Department export-control directive ordered access to Anthropic's two most capable models, Claude Fable 5 and Mythos 5, blocked for any foreign national, a category that includes green-card holders, visa workers, and a company's own non-citizen employees. With no way to verify citizenship in real time at the API layer, Anthropic disabled both models for everyone. Every other Claude model kept running. On June 30 the Commerce Department lifted the order; Fable 5 returned globally on July 1, and Fable 5.1 has since taken over the top of the coding leaderboards.

Nineteen days, then business as usual. That is exactly why it matters. The most capable model in the world went uncallable overnight for reasons that had nothing to do with anyone's work, and it came back on a timeline nobody outside two organizations controlled. Teams that had built on it fell back to older models or stalled. Models people ran on their own hardware never noticed, because there was nothing to switch off. We covered the episode in our guides to what Fable 5 and Mythos 5 are and how the Fable 5 suspension played out.

That is the throughline of everything we cover: whoever controls the infrastructure controls the experience. A capable open model on hardware you own never changes its terms mid-project, never hits a usage cliff, and never goes dark over a policy you had no part in. The frontier gap is real and we will keep being straight about it, but as of September a model that runs on one used GPU is inside the frontier pack on everyday coding, and it stays yours.

So Which One Should You Pick?

For most readers, Qwen3.8-27B: Apache 2.0, multimodal, and the strongest coding score any single-GPU open model has posted. If your agent needs to be fast more than it needs to be smart, Nemotron 3.5 Lightning. If tool-call reliability and a plain license matter more than benchmark rank, Muse Glimmer. On 16 GB, Gemma 4 26B MoE. On a 128 GB workstation with frontier-class ambitions, DeepSeek V4-Flash. If you are fine with hosted access and want weights you can self-host later, GLM-5.3 and DeepSeek V4-Pro give frontier-adjacent results at a fraction of closed-model token cost under MIT terms, and Qwen3.8-Max does the same if you are under the revenue line. For the specific machine to buy or the exact quantized file that fits, our VRAM guide picks up where this page leaves off.

Frequently Asked Questions

What is the best open-source LLM right now?

Qwen3.8-27B for anyone with a 24 GB GPU or a 32 GB Mac: Apache 2.0, reads images and video, and reports 61.7 percent on SWE-bench Pro, the best score any single-GPU open model has posted. If you have data-center hardware, GLM-5.3 and Qwen3.8-2.4T lead the open field. Under 16 GB, Gemma 4.

Is open-source AI as good as ChatGPT or Claude?

On everyday coding and agent tasks, the best open weights now score above GPT-5.6 and Claude Sonnet 5 on vendor-reported SWE-bench Pro numbers, and a single-GPU model scores above GPT-5.5. Anthropic's Fable 5.1, Opus 5, and Opus 4.8 remain clearly ahead, and the closed frontier still leads on the hardest long-horizon reasoning. For most daily work the gap is small enough that hardware and licensing decide the choice.

Can I use an open-source LLM commercially?

Under Apache 2.0, MIT, or OpenMDW-1.1, yes, with attribution and no other conditions; that covers Gemma 4, Qwen3.8-27B, Muse Glimmer, Nemotron, DeepSeek V4, and the GLM family. The Qwen3.8-2.4T, Kimi K3, and LFM licenses are free until your business passes a revenue threshold. Llama 4 excludes EU users. Read the LICENSE file in the repository before shipping.

Can I run Qwen3.8-Max locally?

Technically yes, practically no. The weights are downloadable as Qwen3.8-2.4T-A95B, but 2.4 trillion parameters need a multi-node GPU cluster, and the open checkpoint is text-only with thinking permanently on, unlike the hosted Max version. The self-hostable Qwen for normal hardware is Qwen3.8-27B.

Do I need a GPU to run an open-source LLM in 2026?

No. Apple Silicon Macs with 16 GB or more run quantized small models at conversational speed, and with 32 GB they handle Qwen3.8-27B and Muse Glimmer at Q4. Modern AMD mini PCs with 32 GB of DDR5 run Gemma 4 26B MoE and Nemotron 3.5 Lightning with no discrete GPU. A GPU makes the dense 27B to 30B models feel much faster, but it is no longer required to start.

Did the Fable 5 suspension change these recommendations?

For local inference, no: the picks are set by memory and license. What it changed is the argument. Fable 5 was unavailable to everyone from June 12 to July 1, 2026, because of a government order, with no warning and no recourse for ordinary users, while models people run themselves were unaffected. Access is restored, but the demonstration stands.