Apache 2.0 vs Llama licence: what Indian enterprises need to know.

Most people think "open-source" means no licence problem. They're right — for hobbyists running a model on their laptop. The moment you ship a paid product that includes a model as part of the deliverable, you are redistributing that model, and every open-source LLM licence in the market has an opinion about what that means. Several of those opinions are expensive.

The public conversation about LLM licences is almost entirely written for individual researchers and weekend hackers. What it almost never covers is the vendor's position: the system integrator, the fine-tuning shop, the platform builder who sells a working deployment to an enterprise customer and invoices for it. For that party — which is exactly what Reyatech Systems is — the licence is a procurement item that the customer's legal team will scrutinise, that their cyber-insurance underwriter may query, and that, if chosen carelessly, can expose both vendor and customer to claims that would not otherwise exist.

The two questions every Indian procurement team should ask

Before any model is shortlisted, ask two questions. First: does the licence permit commercial redistribution of fine-tuned weights to a paying third party? Second: what obligations flow down to us as the recipient? The answer to the first eliminates roughly half the models that get casually recommended. The answer to the second determines your insurance exposure, your branding choices, and — in the pharma context — whether you have a latent AUP liability sitting inside your deployment.

Apache 2.0 / MIT — the clean default

Apache 2.0 and MIT are functionally equivalent for redistribution purposes. Both grant a perpetual, irrevocable, worldwide licence to use, reproduce, distribute, and sublicense — including as part of a commercial product — with no field-of-use restriction, no acceptable-use policy, no branding mandate, and no MAU threshold. The only operative obligation is preserving the copyright and NOTICE file in distributed artefacts. For a vendor shipping fine-tuned weights to an enterprise customer on their own hardware, that means one text file in the deliverable package. That is the entire compliance footprint.

Apache 2.0 adds a patent grant that MIT lacks — §3 grants a perpetual, irrevocable, royalty-free patent licence for the contributions in the release. For enterprise customers worried about downstream IP indemnity, this matters: an Apache-licensed base model at least does not create a new patent-infringement exposure vector.

Qwen 3 is the cleanest example in the current market. All open-weight sizes — 0.6B through 32B dense, and the MoE 30B-A3B / 235B-A22B variants — are Apache 2.0. For an Indian pharma vendor shipping a regulated deployment, this means the base-model licence question is resolved at the NOTICE file level, and every clause in the contract with the customer can focus on what actually matters: data governance, CSV obligations, and audit trail requirements.

Llama 3.x — usable but encumbered

Llama 3.1 and Llama 3.3 are governed by functionally identical licences (Meta Llama Community License) that grant redistribution rights — but attach three concrete obligations that create real-world cost for any Indian B2B vendor.

The branding obligation is the one most people skip past. Llama 3 Community Licence §1.b.i.B requires that you "prominently display 'Built with Llama' on the related website, user interface, blogpost, about page, or product documentation." §1.b.i requires that "if you use the Llama Materials to create, train, fine tune, or otherwise improve an AI model... you will name that model in a manner that includes 'Llama' as a prefix." For an Indian CISO building a GMP document assistant that they will present to an FDA auditor, having Meta's brand on their internal AI system is at minimum an awkward conversation and at maximum a procurement objection from the customer's legal team.

The AUP §2.c risk is more subtle but more dangerous for pharma. The Llama AUP — incorporated by reference into every Llama licence — §2.c bans use "related to... regulated/controlled substances." The clause does not define "related to." A hostile reading can reach lawful Schedule II/III therapeutic pharmaceutical work at a company like Dr. Reddy's or Cipla. With Apache 2.0 base models, there is no AUP to accept and nothing to audit.

Reyatech will deploy a Llama-based system if a customer specifically requests it. That request requires a written acknowledgement from the customer of the AUP §2.c interpretive risk and of the "Built with Llama" attribution requirement. This is documented in ADR-002 and is reflected in our standard SOW. See how our deployment architecture handles this.

Mistral MRL — effectively unusable

This one is not a grey area. The Mistral Research License §3.2 defines "Research Purposes" and explicitly excludes "usage by a company's contractor on a revenue-generating activity" and "any distribution by a commercial entity." The Mistral Non-Production License §3.2 prohibits supply "in the course of a commercial activity, whether in return for payment or free of charge."

The models governed by MRL are Mistral Large 2 (versions 2407 and 2411), Pixtral Large, and Ministral 8B. The model governed by MNPL is Codestral 22B v0.1. Selling a fine-tune of any of these models to a paying customer is a direct, unambiguous licence breach. Not an interpretive risk. A breach. This is the primary reason Mistral's commercial models do not appear anywhere in Reyatech's shortlist.

The confusion arises because Mistral also publishes a substantial Apache 2.0 line. A procurement team that approves "a Mistral model" without specifying which one has a meaningful probability of approving a non-redistributable model.

Gemma — high-friction

Section §3.1(a) of the Gemma Terms requires that the vendor's contract with their customer contractually flow down the Gemma Prohibited Use Policy as enforceable terms. Not as a reference. As enforceable contractual terms. This means Reyatech's MSA with Cipla must contain language that makes Google's PUP directly enforceable — and that the customer's legal team must review and accept Google's PUP as part of the vendor contract negotiation.

Section §4 of the Gemma Terms states: "Google reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement." For a deployment on the customer's own hardware — which is precisely the architecture Reyatech sells — the idea that the model vendor retains a remote kill-switch is a structural problem. RBI Outsourcing Master Direction continuity requirements create a genuine legal conflict that no amount of contract drafting can fully resolve.

The trap: families that mix licences

Two common recommendations carry a licence trap that is invisible unless you check the specific model card rather than the family name.

Qwen 2.5-3B. Alibaba's official blog confirmed at the Qwen 2.5 launch that all sizes are Apache 2.0 except 3B and 72B. Qwen 2.5-3B sits under the Qwen Research License, which is non-commercial. A developer who pulls the 3B for a quick PoC, finds it works well, and ships it into a paid product has breached the licence — without ever intending to. Qwen 3 supersedes this problem: Apache 2.0 across all open-weight sizes.

DeepSeek-R1-Distill variants. DeepSeek-R1 ships under canonical MIT. But the R1 release includes distilled variants built on different base models, and those variants inherit the upstream licence of the base, not the MIT wrapper. DeepSeek-R1-Distill-Llama-8B and DeepSeek-R1-Distill-Llama-70B inherit the full Llama 3.1/3.3 Community Licence — including "Built with Llama" branding, the "Llama-" name prefix requirement, the AUP, and the 700M MAU clause. DeepSeek-R1-Distill-Qwen-32B stays under Apache 2.0.

Bottom line.

The Indian enterprise procurement teams we talk to usually treat model licences as a vendor problem. That framing is wrong for on-premise deployments where the fine-tuned weights are part of the delivered artefact. In that structure, the model licence flows through to the customer's contract, and any encumbrance on the model becomes an encumbrance on the deployment. "Built with Llama" on your BMR reconciliation tool is a question you will be asked by your legal team, your auditor, and eventually your regulator — not your AI vendor's legal team.

The practical implication for indemnity and insurance is material. With an Apache 2.0 or MIT base model, a customer purchasing from Reyatech has no Acceptable Use Policy exposure to indemnify against, because there is no AUP. Cyber-insurance underwriters pricing an AI deployment will eventually learn to ask which base model is inside the product. When they do, "Apache 2.0" will be a materially different underwriting answer than "Llama Community Licence with a §2.c pharma carve-out representation."

Pin the SKU, not the brand. Book a discovery call to walk through your base-model choice with our legal counsel.