Best Open-Source LLM in 2026

Eight Apache 2.0 models from 0.6B to 235B with one chat template: the family that turns model choice into a sizing decision.

Qwen 3 wins because choosing an open-source LLM for production is a portfolio decision, not a single-model bet. The family ships eight models under plain Apache 2.0: six dense checkpoints from 0.6B to 32B and two Mixture-of-Experts models, topped by Qwen3-235B-A22B (235B total, 22B active, 128 experts with 8 active per token). Pretraining spans roughly 36 trillion tokens across 119 languages and dialects, double the prior generation. The signature feature is hybrid thinking: one checkpoint reasons step-by-step when thinking mode is on and answers immediately when it is off, so inference-time reasoning cost becomes a per-request dial. The family has also kept moving: the 2507 refresh of the flagship ships as a dedicated non-thinking instruct checkpoint with native 262,144-token context, up from 32,768 native on the original. What this buys in practice is continuity: prototype on the 4B, ship the 32B, scale to the MoE flagship without changing license review, chat template, or serving stack. Its honest limits: the family is text-only, and the flagship is no longer the single most capable open model on vendor coding benchmarks; GLM-5.2 is.

The capability frontier of open weights: MIT licensed, 1M-token context, and within a point of Claude Opus 4.8 on the vendor's agentic coding benchmark.

GLM-5.2 is the strongest answer yet to the question every CTO asks quarterly: are we still paying frontier API prices for work an open model can do? Announced June 17, 2026 with weights on Hugging Face under plain MIT, it is a 753B-parameter Mixture-of-Experts model (roughly 40B active per token) with a 1M-token context window, up from GLM-5.1's 200K. Z.ai's published numbers put it within a point of Claude Opus 4.8 on FrontierSWE and at 81.0 on Terminal-Bench 2.1 against Opus 4.8's 85.0, with SWE-bench Pro at 62.1; treat those as vendor claims, but Simon Willison's independent day-after assessment called it probably the most powerful text-only open weights LLM. The architecture earns its keep at long context: an IndexShare scheme shares one indexer across every four sparse attention layers for a claimed 2.9x per-token FLOPs reduction at 1M tokens, and a multi-token-prediction layer claims up to 20% better speculative-decoding acceptance. The costs are equally concrete: about 1.51TB of BF16 weights means multi-node serving or aggressive quantization, and there is no vision input at all.

The proven MIT pair: the workhorse and the reasoning model that made open chain-of-thought real, now the best-understood self-hosting stack in the field.

The V3 and R1 line remains the reference implementation for owning your own reasoning. R1 is a 671B-parameter MoE with 37B active and 128K context under MIT, trained from DeepSeek-V3-Base, and the R1-0528 refresh is the checkpoint to run: DeepSeek's published numbers show AIME 2025 rising from 70.0 to 87.5 and LiveCodeBench from 63.5 to 73.3, with system prompts properly supported and function calling improved. V3.2 (685B) added DeepSeek Sparse Attention for long-context efficiency, and the distill ladder (six models from 1.5B to 70B, the Qwen-based ones under Apache 2.0, plus an R1-0528 distill onto Qwen3-8B) covers hardware budgets normal companies have. Rank three comes with an honest asterisk: this is now the previous generation. DeepSeek's V4 line shipped in April 2026 with its own MIT open weights on Hugging Face (V4-Pro at 1.6T total and 49B active, 1M context), and the first-party API's deepseek-chat and deepseek-reasoner names are scheduled for deprecation on July 24, 2026, already mapping to V4-Flash. That cuts both ways. You are not buying the frontier; you are buying eighteen months of tooling, quantizations, deployment guides, and known failure modes, which is exactly what a production self-hosting decision should weigh.

The edge and single-GPU default: Google's first Apache 2.0 Gemma generation, from phone-class models with native audio to a 31B flagship on one H100.

Gemma 4 owns the deployment slot the giant MoEs above it cannot touch: models that run where the data lives. Launched March 31, 2026 as the first Gemma generation under plain Apache 2.0, the family spans E2B and E4B edge models (effective 2B and 4B footprints, 128K context, native audio input, offline on phones, Raspberry Pi, and Jetson-class hardware per Google), a 26B MoE activating only 3.8B parameters per token, and a 31B dense flagship with 256K context whose unquantized bfloat16 weights Google states fit a single 80GB H100. All sizes process images and video natively, and the family is trained on over 140 languages. The cadence since launch is the strongest signal: Multi-Token Prediction variants landed April 16, 2026, and Gemma 4 12B Unified arrived June 3, 2026 with text, audio, and image input at 256K context, closing the mid-range audio gap. Google's launch claim of the number-three open-model Arena ranking for the 31B is a vendor claim, but the honest positioning does not depend on it: this is the family you shortlist when the constraint is local, and the ceiling you accept is that a 31B dense model is not a 753B frontier MoE.

Europe's open flagship: Apache 2.0, 675B MoE with built-in vision, and the only place frontier scale, a permissive license, and an EU vendor meet.

Mistral Large 3 is the model you evaluate when the first requirement is a jurisdiction, not a benchmark. Released December 2, 2025 as the flagship of the Mistral 3 family, it is a sparse MoE with 675B total and 41B active parameters including a 2.5B vision encoder, a 256K context window, and a plain Apache 2.0 license across the whole family, which also includes Ministral 3 at 14B, 8B, and 3B as a same-license distill ladder. Two release details matter operationally: the weights ship in FP8 with an NVFP4 checkpoint that Mistral states runs on a single 8xA100 or 8xH100 node via vLLM, and Mistral's hosted API prices it at $0.50 per million input tokens and $1.50 per million output tokens, well under Z.ai's hosted rate for GLM-5.2. Mistral's claim of debuting at number two among open non-reasoning models on LMArena is a vendor claim from launch. What keeps it at rank five is the field, not the model: GLM-5.2 out-benchmarks it on published coding evals, Qwen 3 out-ladders it on family range, and its ecosystem is smaller than either. But for EU-sovereignty deployments that need open multimodal capability with a European vendor behind it, nothing else on this list checks all three boxes.

The famous name with the conditional license: capable multimodal MoE weights, ranked last because every alternative above it asks less of your lawyers.

Llama 4 sits last not because the models are weak but because the deal around them is the worst on this page. The April 5, 2025 release brought two natively multimodal MoE models: Scout (109B total, 17B active, 16 experts) with a claimed 10M-token context window and an Int4 fit on a single H100, and Maverick (400B total, 17B active, 128 experts) with a 1M-token context, both distilled from the still-training Behemoth (288B active, roughly 2T total). The instruct models cover 12 languages with an August 2024 knowledge cutoff. The problem is the Llama 4 Community License. It is free for almost everyone, but it is not open source in the OSI sense: a 700M monthly-active-user gate (measured on the release date), a mandatory and prominently displayed Built with Llama attribution on products that contain the materials, a requirement that distributed derivative models carry Llama at the start of their name, and redistribution paperwork. In 2024 those conditions were the price of frontier open weights. In mid-2026, Apache 2.0 and MIT alternatives match or beat Llama 4 across this list, so the conditions need a justification they no longer have. Pick it for the ecosystem or a specific capability fit, and read the license first.