Best open-weight LLMs (2026)
A leaderboard rank tells you almost nothing about running the model under a contract. License terms, serving cost and tool-calling reliability decide that.
The short answer
Qwen is the widest open-weight family, with Apache 2.0 on its smaller models. DeepSeek is the close second under MIT. Check Llama's license with a lawyer before you commit.
A benchmark score describes a model in a lab. Your production deployment adds a license your legal team has to sign off. It adds a serving bill that grows with traffic, and tool-calling that has to survive the fifth retry.
Qwen's newest flagship ships under Qwen's own license, which adds conditions at scale, so the clean grant now depends on the size you pick. OpenAI's gpt-oss is the simplest Apache 2.0 option if your agent leans on tool calls.
We build agent and RAG systems for clients who ask this exact question. Your answer changes with where the model has to run.
Key takeaways
3 things that decide this
- 01Qwen is the best open-weight LLM for most production work when a smaller Apache 2.0 model covers the job; DeepSeek wins when serving cost decides.
- 02We ranked five families on license terms, serving cost, tool-calling and fine-tuning support, read from each vendor's own license and model card.
- 03Before you build, read the license on the exact checkpoint you plan to ship. Terms now differ by model size inside the same family.
How we ranked
Verified
We ranked five open-weight families on four criteria: license terms, serving cost at scale, tool-calling and fine-tuning support. Order reflects how often each one is the safe default for a paying client, not its benchmark rank.
For every entry we read the license on the vendor's own model card or license page. We also checked the architecture, the size and the current release. We read them on 11 November 2026. Where a term could change with a point release, we say so instead of freezing a claim that will go stale.
We left out closed-weight, API-only models. That comparison has its own page: see self-hosted versus managed LLM below.
Hashlogics publishes this page. We build custom software and are not one of the tools ranked here. No company paid for its place.
- License terms
- Whether commercial use is unconditional, or gated behind a usage cap or a use-case restriction.
- Serving cost at scale
- Whether the architecture (dense versus mixture-of-experts) lets you serve it without a large GPU fleet.
- Tool-calling reliability
- Whether structured output and function calls hold up across many turns, past the first try.
- Fine-tuning support
- Whether the ecosystem around the model (adapters, quantized builds, serving frameworks) is mature enough to bet on.
The families compared
License terms as published by each vendor, checked 11 November 2026.
| Family | License | Architecture | Pick it when |
|---|---|---|---|
| Qwen | Apache 2.0 on smaller releases; Qwen's own license on the newest flagship | Dense and mixture-of-experts variants | A smaller Apache 2.0 model covers the job |
| DeepSeek | MIT on current V4 weights | Mixture-of-experts | Serving cost matters more than raw score |
| gpt-oss | Apache 2.0 | Mixture-of-experts, two sizes | Your agent depends on tool calls and you want one clean license |
| Llama | Llama 4 Community License, usage-gated | Mixture-of-experts on Llama 4 | You are under the 700 million user cap and outside the EU restriction |
| Mistral | Apache 2.0 on Small, Ministral and Large 3; Mistral's own license on Large 4 | Dense and mixture-of-experts | You want a European vendor and a smaller model that fine-tunes well |
The ranking
Ordered by how often each is the safe default for a paying client, not by benchmark position.
- 01
Qwen ↗
Apache 2.0 on smaller models, widest size range
Qwen's smaller releases ship under Apache 2.0, a real open-source license. Qwen3.8-27B, released in August 2026, is one of them. The flagship from the same release carries Qwen's own license instead.
That license is still broad, but it adds conditions. Your product must show the model name once it passes 100 million monthly users or a stated revenue line. A model-as-a-service or AI-assistant business above a revenue threshold needs a separate license from Qwen.
So your license conversation depends on which size you pick. Before you commit to a new checkpoint, confirm your serving framework supports it.
Best for
- Teams whose model size fits the Apache 2.0 releases, with no lawyer follow-up
- Deployments that want one family from laptop-sized to cluster-sized models
Not for
- Model-as-a-service businesses planning to resell the flagship without a separate license
- Anyone who needs a long enterprise support contract from the vendor
- License
- Apache 2.0 / Qwen's own, by size
- Sizes
- Dense and MoE, small through flagship
MIT license, mixture-of-experts, cheap to serve
DeepSeek's current V4 weights carry an MIT license, which is about as permissive as licensing gets. You can use, change and redistribute them commercially with no royalty and no cap.
Its mixture-of-experts design only activates a fraction of total parameters per token. V4-Pro has about 1.6 trillion parameters with 49 billion active, and a smaller V4-Flash sits beside it. That's the whole argument for choosing it: lower inference cost than a dense model of similar quality.
Check the specific checkpoint before you commit. DeepSeek released some older weights under a separate, stricter license. The MIT grant also covers the weights only, not the hosted API service.
Best for
- Self-hosted deployments where inference cost per token is the deciding number
- Teams that already run mixture-of-experts serving infrastructure
Not for
- Anyone assuming every DeepSeek checkpoint carries the same license without checking
- Teams wanting a single dense model with a simpler serving story
- License
- MIT on current V4 weights
- Architecture
- Mixture-of-experts
OpenAI's Apache 2.0 models, built for tool calls
OpenAI's gpt-oss models ship under Apache 2.0, with no user cap and no use-case clause. The family has two sizes: gpt-oss-120b and gpt-oss-20b, both released in August 2025.
Both are mixture-of-experts. The 120b model has about 117 billion parameters with 5.1 billion active, and its quantized weights fit on a single 80GB GPU. Function calling and structured outputs are part of the model card, which matters if your agent makes many tool calls.
Your stack has to speak OpenAI's harmony response format, and the model card says it won't work correctly without it. It's also smaller than the frontier flagships from Qwen, DeepSeek and Mistral.
Best for
- Agents that lean on function calls and structured output
- Teams that want one clean license and a single-GPU serving story
Not for
- Serving stacks that cannot apply the harmony response format
- Workloads that need a frontier-sized model
- License
- Apache 2.0
- Sizes
- 120b and 20b, MoE
- 04
Llama ↗
Deep ecosystem, license reads with a lawyer
Llama has one of the deepest tooling ecosystems of any open-weight family. Serving frameworks, fine-tuning libraries and quantized builds all support it, and that maturity is why many teams still start here.
The Llama 4 Community License is not Apache or MIT. Companies above 700 million monthly active users must ask Meta for a license. Any model you improve with Llama has to carry 'Llama' at the start of its name. Meta has also excluded EU-based companies from the Llama 4 multimodal models.
None of that rules Llama out. It means the license is a real step in your diligence. Meta's newer Muse Glimmer, a 30B model released in August 2026, ships under Apache 2.0, so check which Meta model you're actually picking up.
Best for
- Teams that value a wide tooling and fine-tuning ecosystem
- Deployments well under the user cap and outside the EU restriction
Not for
- EU-based companies building on the Llama 4 multimodal models
- Anyone who needs weights under a standard OSI license for a compliance policy
- License
- Llama 4 Community License
- Ecosystem
- Wide tooling support
Apache 2.0 on most of the open line, strong fine-tuning fit
Mistral's Small, Ministral and Large 3 models ship under Apache 2.0, the same clean grant as Qwen's smaller models. Large 3 is now marked deprecated, with retirement set for April 2027.
Its replacement, Mistral Large 4, has open weights under Mistral's own Ecosystem Open-Weight license. That license allows commercial use but asks model-as-a-service businesses to get a separate license first. Mistral's OCR and embedding models are API products.
Its smaller open models fine-tune well on modest hardware. That matters if you want a model trained on your own documents, not a general-purpose giant.
Best for
- Teams fine-tuning a smaller model on their own data
- A European vendor requirement alongside a clean open license
Not for
- Teams that need Mistral's OCR or embedding models as open weights
- Model-as-a-service businesses planning to resell Large 4 without a separate license
- License
- Apache 2.0 / Mistral's own on Large 4
- Strength
- Fine-tuning on smaller checkpoints
- LicenseRead the actual grant, not a summary site.
- ArchitectureMixture-of-experts serves cheaper at scale.
- Tool-callingTest the fifth retry, not the first.
- Serving stackConfirm your framework supports the checkpoint.
The benchmark score answers none of these four questions.
“I am extremely happy with the results and would highly recommend Hashlogics to anyone.”
Daniel Khin · CEO, PremiumAudit.io
Choosing between open weights and a managed API?
Send us your expected volume and the license constraints you are working under. We will tell you which model family actually clears them. Scoping costs nothing.
When an open-weight model is the wrong call
Self-hosting an open-weight model trades your per-token bill for infrastructure you now own. That means GPUs, scaling, upgrades and the on-call rotation when inference goes down at 2am. That trade only pays off at steady volume.
A managed API is usually the right starting point for a new product. You can revisit open weights once traffic and cost patterns are real.
When none of these fits, we build the agent on a managed model and keep it swappable, so you can move later without a rewrite. See our AI agent development work.
- 01If request volume is low or unpredictable, a managed API's per-token pricing beats owning idle GPU capacity.
- 02If your team has no one who wants to own model upgrades and serving infrastructure, that cost does not disappear by choosing open weights.
- 03If the license restricts your region or your user count, confirm that before building, not after.
Questions teams ask
01What is the best open-weight LLM?+
Qwen is the best open-weight LLM for most production work when one of its Apache 2.0 sizes covers the job. DeepSeek is the pick when serving cost decides, and gpt-oss when your agent leans on tool calls.
02Is Qwen or Llama better for a commercial product?+
Qwen is the safer pick if a smaller Apache 2.0 model covers your job. Its newest flagship carries Qwen's own license, which adds conditions at scale. Llama's license gates use above 700 million monthly users and excludes EU companies from its multimodal models.
03Is DeepSeek safe to use commercially?+
Yes, for current-generation weights. DeepSeek V4 carries an MIT license, which permits commercial use without royalties. Confirm the checkpoint you plan to use, since some older DeepSeek weights shipped under a stricter license.
04How were these models ranked?+
We ranked five families on license terms, serving cost, tool-calling and fine-tuning support, checked against vendor model cards on 11 November 2026. Hashlogics publishes this page and is not one of the tools ranked; no company paid for its place.
05How do I compare open-weight LLMs fairly?+
Run each model on the same set of your real tasks, served the same way, with the same prompts. Count failed tool calls across many turns, then read the license on the exact checkpoint before you compare cost.
06Is gpt-oss good enough for production?+
Yes, for agents and tool-heavy work that fit its size. It's Apache 2.0 and runs on a single 80GB GPU, but your serving stack must use OpenAI's harmony format, and it's smaller than the frontier flagships.
07Does open-weight mean the same thing as open-source?+
Not exactly. A license on the weights tells you what you can do with the model file. It does not disclose the training data or the process used to produce it. Apache 2.0 and MIT weights sit closest to standard open source; a custom license like Llama's sits further from it.
08Why does the mixture-of-experts architecture matter for cost?+
A mixture-of-experts model activates only a fraction of its total parameters for any given token. Inference costs less than a dense model of similar quality. DeepSeek and some Qwen variants use this design, part of why they serve cheaply at scale.
09Should we self-host instead of using a managed LLM API?+
Only once volume and cost patterns are real. Self-hosting removes the per-token bill but adds GPU infrastructure, scaling and on-call ownership. A managed API is the safer default until traffic makes the tradeoff worth revisiting.
10Why isn't Gemma or Phi on this list?+
Both are real options outside this four-family ranking. Google's latest Gemma generation ships under Apache 2.0, while earlier Gemma models keep Google's custom terms — so check which generation you are picking up. Microsoft's Phi targets small, cheap-to-run models rather than competing at the frontier, which is a different decision than the one this page ranks. Bring either to a scoping call once your license and serving constraints are set.
Related reading
- Self-hosted vs managed LLM →The infrastructure tradeoff behind this ranking, in full.
- Model deprecation risk →What happens when the checkpoint you built on goes away.
- Model engineering services →Fine-tuning and deployment work on top of a chosen model.
- AI agent development →Where the model choice on this page actually gets used.
- Best LLM inference providers →Once you've picked a model, who serves it fast enough to ship.
