Best Open-Weight LLM: MiniMax M3 Leads

The best open-weight LLM right now is MiniMax M3 (69.8, Supported). GLM-5.1 and GLM-5 follow. The previous DeepSeek V4 Max and Kimi-led ordering came from a narrower verified lane; BenchAlign v5 ranks the full comparable evidence set and labels thinner positions instead of hiding them.

That's a significant shift. Two years ago, Llama dominated the open source conversation. Today, Chinese labs — DeepSeek, Moonshot AI, Zhipu AI, and Alibaba — hold most of the top positions among open weight models. The best open source LLMs in 2026 are not where most people expect them to be.

Top open-weight LLMs ranked

Table 1 Download CSV

Rank	Model	Type	License	Score	Evidence
1	MiniMax M3	Non-Reasoning	Open Weight	69.8	Supported
2	GLM-5.1	Reasoning	Open Weight	67.7	Supported
3	Inkling	Hybrid	Open Weight	67.5	Supported
4	GLM-5	Non-Reasoning	Open Weight	66.1	Supported
5	MiniMax M2.7	Non-Reasoning	Open Weight	64.1	Supported
6	GLM-5.2	Reasoning	Open Weight	64	Estimated
7	GLM-4.7	Reasoning	Open Weight	61.2	Supported
8	Gemma 4 31B	Reasoning	Open Weight	61.1	Supported
9	Qwen3.5-27B	Reasoning	Open Weight	60.7	Supported
10	DeepSeek V4 Pro	Non-Reasoning	Open Weight	60.7	Supported

The score sets the order. Supported and Estimated describe how much direct, independent evidence backs each position.

This table reveals something non-obvious: the models with the loudest individual benchmark rows are not always the strongest overall. MiniMax M3 wins on breadth and calibrated evidence. A DeepSeek or GLM variant can still be the better choice for one task, but one exceptional coding result no longer carries the whole model. The matching open-weight ranking page now reads from this same public lane; this guide owns the deployment decision, not a second ordering.

How close are open source models to proprietary ones?

The honest answer: competitive on individual tasks, still behind overall. These two tables are regenerated from the same BenchAlign v5 artifact, so the comparison cannot drift into separate ranking lanes.

Overall frontier

Table 2 Download CSV

Rank	Model	Type	License	Score	Evidence
1	Claude Mythos 5	Reasoning	Proprietary	83.9	Supported
2	Claude Fable 5	Reasoning	Proprietary	83.7	Supported
3	GPT-5.6 Sol	Reasoning	Proprietary	82	Supported
4	Kimi K3	Reasoning	Pending	81	Supported
5	Claude Opus 4.8	Reasoning	Proprietary	78.3	Supported

Open-weight frontier

Table 3 Download CSV

Rank	Model	Type	License	Score	Evidence
1	MiniMax M3	Non-Reasoning	Open Weight	69.8	Supported
2	GLM-5.1	Reasoning	Open Weight	67.7	Supported
3	Inkling	Hybrid	Open Weight	67.5	Supported
4	GLM-5	Non-Reasoning	Open Weight	66.1	Supported
5	MiniMax M2.7	Non-Reasoning	Open Weight	64.1	Supported

The overall gap between MiniMax M3 at 69.8 and Claude Mythos 5 at 83.9 is 14.1 points. Fable 5, the generally available Claude 5 row, is almost as high at 83.6. Open weights remain compelling for control and price, but the current score does not support calling the capability race tied.

Where open source models already match or beat proprietary ones:

Table 4 Download CSV

Category	Open-weight leader	The receipts
Math	GLM-5 (Reasoning)	AIME 2025: 98 · HMMT 2025: 95 — competitive with the best proprietary math rows
Knowledge	GLM-5 (Reasoning)	MMLU 96 · GPQA 94 · SuperGPQA 92
Competitive coding	DeepSeek V4 Pro (Max)	LiveCodeBench 93.5 (Kimi K2.6: 89.6), both ahead of Claude Opus 4.6 at 76 on that benchmark
Multilingual	GLM-4.7	MGSM 94, ahead of most proprietary models

Where the gap remains wide:

Software engineering is the clearest gap: the best open-weight SWE-bench Verified score is 80.6 (DeepSeek V4 Pro (Max)) against GPT-5.3 Codex's 85, and on the harder SWE-bench Pro open models top out around 67 versus 90. Agentic tasks show the same pattern, with open models trailing significantly on BrowseComp, TerminalBench, and OSWorld. The subtler gap is consistency: proprietary models perform well across all categories simultaneously, while open models tend to spike on specific benchmarks and dip on others.

Best open source LLM by use case

Best for math and reasoning

GLM-5 (Reasoning) is the clear winner. AIME 2025: 98. HMMT 2025: 95. BRUMO 2025: 96. Math500: 92. These are near-perfect scores on graduate-level competition math. No other open weight model comes close.

Runner-up: Step 3.5 Flash (AIME 2025: 97.3) and GLM-4.7 (HMMT 2025: 97.1) are both strong math alternatives with lower overall resource requirements.

Best for coding

This depends on which coding benchmark you care about:

Table 5 Download CSV

Benchmark	Open-weight leaders
HumanEval (function generation)	Kimi K2.5 at 99 — saturated, no longer differentiates
LiveCodeBench (competitive programming)	DeepSeek V4 Pro (Max) 93.5, DeepSeek V4 Flash (Max) 91.6, Kimi K2.6 89.6
SWE-bench Verified (real bug fixing)	GLM-5 77.8, Kimi K2.5 76.8, GLM-4.7 73.8
SWE-bench Pro (harder engineering)	GLM-5 (Reasoning) 67 — still well below GPT-5.3 Codex (90)

On the calibrated coding score, GLM-5.2 leads the open-weight field at 62.3 with an Estimated label. Kimi K2.7 Code follows at 61.9 with Supported evidence. If one raw benchmark matters most, use that benchmark page; if you need a production default, the evidence label is a better tiebreaker than half a score point.

→ See the full coding leaderboard · SWE-bench Pro explained

Best for knowledge and question answering

GLM-5 (Reasoning) dominates knowledge benchmarks: MMLU 96, GPQA 94, SuperGPQA 92, SimpleQA 92. The non-reasoning GLM-5 variant is nearly as strong at MMLU 91.7 and SimpleQA 84.

Qwen3.5 397B (Reasoning) is a solid second choice with MMLU 91, GPQA 89, and more balanced performance across categories.

For factual accuracy specifically, check SimpleQA scores — they measure whether models hallucinate less. GLM-5 (Reasoning) leads open models at 92.

Best for self-hosting and fine-tuning

Self-hosting economics favor smaller models. Running a 397B-parameter model requires multiple high-end GPUs. Here's where the practical sweet spot sits:

• Mistral Small 4 (24B parameters, 256K context): Not currently public-ranked. It fits on a single consumer GPU with quantization, and Mistral's Apache 2.0 license is genuinely permissive for commercial use
• Gemma 4 31B (31B parameters, 256K context): Not currently public-ranked. Google's latest open model fits on a single high-end consumer GPU and offers strong LiveCodeBench (80) performance
• DeepSeek Coder 2.0: 50 overall, but $0.27/$1.10 via API makes it cheaper than self-hosting for most teams. Self-host only if data privacy requires it

For fine-tuning specifically, Mistral and Qwen models have the most mature fine-tuning ecosystems with well-documented tooling.

Best for privacy-sensitive deployments

If data cannot leave your infrastructure, your options are:

1. MiniMax M3, GLM-5.1, GLM-5, or MiniMax M2.7 for maximum measured capability (large rows still require serious GPU infrastructure)
2. Mistral Small 4 for the best quality-to-resource ratio (runs on a single A100 or equivalent)
3. Llama 3.1 405B for the broadest ecosystem support and community tooling, despite lower benchmark scores (43 overall)

All open weight models can be self-hosted with no API dependency. The real constraint is GPU cost: running a 400B+ model costs $2-5K/month in cloud GPU compute, which only makes sense above roughly 50M tokens per month versus API pricing.

Best for cost-sensitive applications

If cost per token drives your decision, these are the open weight models available through low-cost APIs:

Table 6 Download CSV

Model	API Price (input/output per 1M tokens)	Overall Score
DeepSeek Coder 2.0	—	50
DeepSeek V3.2 (Thinking)	~$0.55 / $2.19	57.9
Kimi K2.5	$0.60 / $3.00	59.6

For comparison, GPT-5.4 Pro costs $30/$180 and Claude Opus 4.6 costs $5/$25. DeepSeek Coder 2.0 remains a cheap open-weight option to self-host, but we no longer attach a current hosted API token price to the exact row without a first-party source.

Lab by lab: DeepSeek, Qwen, GLM, Llama

DeepSeek

DeepSeek has the strongest brand recognition among open-weight LLMs, but its canonical V4 Pro row now sits below MiniMax M3 and the leading GLM rows on the overall table. DeepSeek's advantage remains pricing and deployment flexibility. The Max and High configurations retain striking raw coding results, but they do not receive separate overall v5 positions without enough comparable evidence.

Qwen (Alibaba)

Alibaba's open-weight strength is breadth rather than one dominant flagship. Qwen3.5-27B and Qwen3.5-122B-A10B both sit around 60.5 with Supported evidence, while Qwen3.5 397B (Reasoning) lands at 59.2 with an Estimated label. Qwen remains one of the easiest families to size down and deploy across hardware budgets.

GLM (Zhipu AI)

Zhipu AI's GLM family is the strongest open-weight cluster in the current top tier. GLM-5.1 scores 67.8, GLM-5 scores 66.0, and GLM-5.2 scores 64.0 with an Estimated label. GLM-4.7 remains a balanced lower-cost option around 61.

Llama (Meta)

Llama's position has changed dramatically. Llama 4 Scout scores 39.7 and Maverick 23.4, both far below the leading MiniMax and GLM rows. Meta's models remain relevant for ecosystem support, community tooling, and broad cloud availability. On current measured capability, they are not competitive with the open-weight frontier.

Mistral

Mistral Small 4 is not currently public-ranked. Its strength remains efficiency, a 256K context window, and European deployment options for organizations with sovereignty constraints.

NVIDIA Nemotron

NVIDIA's Nemotron 3 Ultra is now tracked as a 550B total / 55B active open-weight model with a 1M context window, but we keep its score marked as coming soon until the full public benchmark set is available. Nemotron 3 Super 120B A12B and Super 100B remain the practical deployed rows with current benchmark coverage. NVIDIA's integration with its own GPU tooling gives Nemotron models a deployment advantage on NVIDIA hardware.

What "open source" actually means for LLMs

Not all "open weight" models are truly open source. The distinction matters:

Open weight means the model weights are downloadable and you can run inference locally. This is what most people mean when they say "open source LLM." GLM-5, Qwen3.5, DeepSeek, and Mistral models are all open weight.

Open source in the strict OSI definition requires the training data, training code, and model weights to all be available. Almost no frontier LLM meets this bar. OLMo from AI2 is one of the few that does.

Permissive license vs restricted license is often more important than the open-weight distinction. Mistral uses Apache 2.0 (fully permissive for commercial use). Meta's Llama license restricts use above 700M monthly active users. Many Chinese open weight models have custom licenses that may restrict commercial deployment in certain jurisdictions.

For most teams, the practical question is: "Can I download the weights, run the model on my infrastructure, and use it in my product without paying per-token fees?" For all models on this list, the answer is yes — but read the license before deploying in production.

How we rank open source models

Open-weight and proprietary models use the same BenchAlign v5 contract. The scorer calibrates benchmark protocols and standardized signals onto comparable scales, then pools partial evidence without converting missing cells into failures. Supported and Estimated are evidence states, not separate ranking lanes.

Provider-reported evidence can inform the estimate, but it cannot earn a Supported label by itself. The full evidence contract is documented on the methodology page.

→ See the full open source leaderboard · Compare all models · Best LLM for coding

---