What is the difference between open weights and open source AI?

Open weights means the trained parameters are downloadable and runnable, which is what almost every so-called open AI model actually offers. Open source in the strict sense would additionally require the training code, training data and full reproducibility. Nearly no frontier-class model meets that stricter bar, so open weights is the accurate term for GLM-5.3-Flash, Muse Glimmer 30B and DeepSeek V4.

Which 2026 AI models have genuinely permissive licenses?

GLM-5.3-Flash ships MIT-licensed weights, DeepSeek V4 ships MIT, and Muse Glimmer 30B ships Apache 2.0 with no usage caps and no revenue thresholds. Qwen3.8-27B is Apache 2.0. These impose no field-of-use restrictions, no revenue triggers, and no acceptable-use addenda that a lawyer needs to review before commercial deployment.

Is Apache 2.0 better than MIT for AI models?

For AI specifically, Apache 2.0 has one meaningful advantage: an express patent grant and a patent-retaliation clause. MIT is silent on patents. Both permit commercial use, modification and redistribution with attribution. If your legal team cares about patent exposure from model architecture claims, Apache 2.0 is the safer default.

Does a permissive license matter if I will never self-host?

Yes, as leverage rather than as deployment. Permissively licensed weights are a credible exit from a vendor, which changes contract negotiations and protects you from deprecation, since a vendor cannot retire a model you hold. Many teams license weights they never run, purely to remove single-vendor risk from their architecture.

Quick Answer

Open Weights vs Open Source vs Source-Available AI

Published: August 27, 2026

The Short Answer

Three terms get used interchangeably in 2026 and mean quite different things:

Term	What you get	Can you deploy commercially?
Open source (strict)	Weights + training code + data + reproducibility	Yes — but almost nothing qualifies
Open weights	Downloadable, runnable parameters	Depends entirely on the license attached
Source-available	Downloadable with restrictions	Conditionally, with legal review

Nearly every model marketed as “open source AI” is actually open weights. The useful question is never “is it open?” — it is “which license, and what does that license forbid?”

Where 2026’s Models Actually Land

Verified against model cards and vendor announcements, August 27, 2026:

Model	Released	License	Restrictions
GLM-5.3-Flash (Z.ai)	Aug 26, 2026	MIT	None material
Muse Glimmer 30B (Meta)	Aug 10, 2026	Apache 2.0	No usage caps, no revenue thresholds
Qwen3.8-27B (Alibaba)	Aug 14, 2026	Apache 2.0	None material
DeepSeek V4	2026	MIT	None material
Qwen3.8-Flash-Next	Aug 26, 2026	Open weights — check model card	Verify before commercial use
GLM-5.3 (Z.ai)	Aug 14, 2026	Staged, not published at launch	API-only at launch
Gemini 3.7 Flash	Aug 13, 2026	Proprietary	API-only
Claude Opus 5, GPT-5.6 Sol	2026	Proprietary	API-only

Two things in that table are worth pausing on.

Meta’s reversal is real. Muse Glimmer 30B shipped Apache 2.0 with no usage caps and no revenue thresholds — an explicit departure from the old Llama license, which gated large deployments behind a monthly-active-user threshold. That threshold was the single most-litigated clause in open AI licensing, and Meta removed it.

Z.ai went both directions in twelve days. GLM-5.3 (August 14) shipped with weights staged rather than published, breaking Z.ai’s open-weight-first reputation. GLM-5.3-Flash (August 26) shipped MIT. Vendor reputation is not a license. Read the card for the specific model, every time.

MIT vs Apache 2.0 — The Only Distinction That Matters

Both permit commercial use, modification, redistribution and sublicensing with attribution. The difference is patents.

	MIT	Apache 2.0
Commercial use	✅	✅
Modify and redistribute	✅	✅
Attribution required	✅	✅
Express patent grant	❌ silent	✅ granted
Patent retaliation clause	❌	✅
Change notice required	❌	✅

Apache 2.0’s patent grant is the AI-specific advantage. Model architectures are patentable, and several 2026 architectures — sparse attention variants, routing schemes, speculative decoding designs — sit on patentable ground. Apache 2.0 gives you an express grant from contributors and terminates that grant if you sue them over it. MIT says nothing, which means you rely on implied license and jurisdiction-specific doctrine.

For most teams this never becomes concrete. For teams with patent counsel or acquisition exposure, it is a real preference for Apache 2.0.

The Restrictions to Look For

When a license is not MIT or Apache 2.0, these are the clauses that actually bite:

Revenue or user thresholds. “Free below X monthly active users” is a licence that expires exactly when your product succeeds. Structurally, it is a deferred sales contract.

Field-of-use restrictions. Bans on military, surveillance, medical or competitive-model-training use. The last one is common and often overlooked: many licenses forbid using outputs to train a competing model, which quietly rules out distillation.

Acceptable-use policies incorporated by reference. The dangerous shape. The AUP is a separate document the vendor can revise unilaterally, so your obligations change without your agreement.

Redistribution and naming requirements. Derivative-naming rules and mandatory attribution strings that are trivial to comply with and easy to violate accidentally.

Staged availability. Not a clause, but the same practical effect. “Weights coming soon” is not a license. GLM-5.3 demonstrated this in August 2026 — plan against what is published, not what is promised.

Why It Matters Even If You Never Self-Host

Most teams that care about licenses will never run the weights. Self-hosting a 320B model needs roughly 192GB of GPU memory and only pays for itself above millions of tokens per hour. That is not most people.

Permissive weights are still worth real money, for three reasons:

1. Deprecation immunity. Vendors retire models. DeepSeek retired the deepseek-chat and deepseek-reasoner endpoint names on July 24, 2026. Weights you hold cannot be turned off.

2. Negotiating leverage. “We can leave” is a materially different position from “we cannot.” A credible migration path changes enterprise pricing conversations even when you never exercise it.

3. Compliance optionality. When a regulator, a customer security review, or a new data-residency rule arrives, having a self-hostable model already validated turns a crisis into a configuration change.

Budget the license as risk reduction, not as a saving. That framing survives audit; “we’ll save money by self-hosting” usually does not.

The Practical Checklist

Before committing to any open-weight model in production:

Read the actual license file in the repo — not the blog post, not the press coverage.
Search for thresholds — MAU, revenue, headcount. Any number in a license is a future invoice.
Check for an incorporated AUP — a separately-revisable document is a moving obligation.
Check the distillation clause — can you train on outputs?
Confirm weights are actually published, not staged.
Record the license and version you downloaded — licenses change between releases of the same family.

If steps 1-6 take more than twenty minutes, the model is not permissively licensed regardless of what the announcement said. MIT and Apache 2.0 take two minutes. That difference is itself the signal.