The Open-Source LLM Guide: Best Open-Weight Models in 2026

“Open source” stopped being a compromise somewhere in the last two years, and the 2026 numbers prove it. The best open-weight models now sit within one SWE-bench point of the most expensive proprietary systems, at a tenth of the API price, with weights you can download, audit, fine-tune, and serve from your own hardware. This guide maps the entire open-weights landscape as we measure it: who leads, what each model is actually good at, what the licenses really say, and when self-hosting beats renting.

Everything below comes from the RankLLMs dataset - the same verified numbers that power the AI leaderboard. All prices are per million tokens.

The 2026 Open-Weights Landscape

Three facts jump out of this table, and each one changes how you should shop.

Fact 1: The Coding Gap Has Closed

DeepSeek-V4-Pro-0813 scores 81.0 on SWE-bench Verified. The best proprietary model, Claude Fable 5, scores 84.5 - and GPT-5.6 Sol scores 81.2, a statistical tie with DeepSeek at more than four times the price. For issue-resolution and code-generation work, asking “can open-source models code yet?” is a 2024 question. The live questions now are about terminal proficiency, cost per solved task, and hosting economics - which we work through in the best LLM for coding guide.

Where the proprietary frontier still pulls away: the hardest multi-step reasoning, live-terminal agent work at the very top (Claude Fable 5 and GPT-5.6 Sol own Terminal Bench’s peak scores), and the newest benchmark formats where training recency matters.

Fact 2: You Have Genuinely Different Options Now

The open-weights field is no longer one model family with cheaper clones. The leaders have distinct personalities:

• Kimi K3 - the open-weights champion. Highest composite score among open models (56.0), an 85.7 Terminal-Bench that embarrasses most proprietary mid-tiers, and a 1M context window. Pick it when you want the closest thing to the closed frontier with self-hosting rights. Full analysis in the Kimi K3 scorecard.
• DeepSeek-V4-Pro-0813 - the economics play. Fastest top-tier model we measure (176 tokens per second), 81.0 SWE-bench, and a family of variants (Pro-Max, Flash-Max) that spans premium to $0.14/M ultra-budget. Pick it for volume coding workloads and API-or-self-hosted flexibility.
• Qwen3.8 Max - the balanced all-rounder. 79.4 SWE-bench at $2.02 flat, strong multilingual coverage inherited from the Qwen lineage, and the most mature fine-tuning ecosystem of any family here.
• GLM-5.3-Flash - the volume weapon. $0.19 per million tokens is not a typo. At 78.2 SWE-bench it runs classification, extraction, code-review triage, and summarization at costs that make always-on pipelines viable. When a task does not need the frontier, this is the default.
• Tencent Hy4 Preview - one to watch. 77.0 SWE-bench in preview form, currently free while Tencent gathers production data. Worth piloting; do not build a business on preview pricing.

Fact 3: “Free Weights” Still Costs Something - Plan For It

Open weights move cost from per-token billing to fixed infrastructure, and the honest math has three tiers:

• API tier (start here). Every model above is available through its provider’s API. You get the open-weights license guarantees (no training on your data under most of these licenses, though verify per provider) without owning any infrastructure. The prices in the table apply.
• Hosted open inference (the middle path). Services serve open weights on demand - you get weight portability across vendors and no hardware ownership, paying a modest premium over raw API pricing.
• Self-hosting (the break-even game). Rough planning numbers: a quantized (4-bit) small model fits a single 24GB GPU; dense models in the 70B class need 48GB+ or multi-GPU; frontier-scale mixture-of-experts systems are datacenter territory. Self-hosting wins when volume is high and sustained, when data cannot leave your network, or when you need fine-tuned behavior no API offers. Below a few hundred million tokens monthly, it usually does not pay.

Licensing is the other cost center. The 2026 open-weight releases mostly permit commercial use, but they are not uniform: some are permissive in the Apache 2.0 / MIT tradition, while others attach acceptable-use conditions or attribution requirements. Two minutes on the model’s official license page (linked from each model scorecard) prevents a legal review surprise later.

Open vs Proprietary: The Five-Question Decision

1. Does your data need to stay on your infrastructure? If yes (healthcare, legal, finance, defense), the decision makes itself.
2. Is your workload high-volume and routine? Volume favors GLM-5.3-Flash or DeepSeek-V4-Flash-Max; routine favors open models with fine-tuning.
3. Do you need the last 3-5 points of SWE-bench? If a failed task costs more than thousands of tokens, the proprietary frontier (Claude Fable 5) pays for itself on hard work.
4. Will you fine-tune? Only open weights can truly be specialized. Proprietary “custom models” are configuration, not training.
5. Can you operate GPUs? Honest answer required. Self-hosting is a DevOps commitment, not a checkbox.

Most production teams land on routing: an open-weights workhorse for the bulk of traffic, a proprietary frontier model for the hard tail. That architecture costs less than either pure approach.

How We Got Here

Two years ago this guide would have been about Llama and Qwen 2.5 trading blows at the 7-8B scale - a comparison we preserved in Qwen 2.5 7B vs Llama 3.1 8B. The open-weights frontier then climbed through the 2025 releases to today’s position, where the top open models challenge for the overall podium, not just the value bracket. The trajectory is the story: every capability gap the proprietary labs opened has closed within quarters, while the price gap has only widened.

Where to Go Next

• AI Model Leaderboard - the current standings, open and proprietary together
• Best LLM for Coding - cost per solved task across the whole market
• Our Methodology - where every number on this page comes from
• Claude 3.5 Sonnet Benchmarks - the model that started the agentic coding era