Falcon goes Apache 2.0: the month a license clause moved a leaderboard
Falcon-40B launched under a TII license that asked for ten percent of revenue above a million dollars. Five days later the license was Apache 2.0, and within the week the model was at the top of the Open LLM Leaderboard with a Hugging Face blog post calling it the first truly open model of its class. On license terms as a competitive lever.
What the first license said
The Falcon-40B repository on Hugging Face got its license file on May 26, and the metadata tag on May 30 read tii-falcon-llm. That license is worth reading now that it is gone. Commercial use of Falcon or any derivative required applying to TII for permission. TII would set a royalty as a percentage of revenue, and unless the grant said otherwise, the percentage was ten. The royalty kicked in once attributable revenue in the preceding twelve months passed one million dollars, and the licensee had to account to TII in writing, once a year, for all revenue attributable to the model. Derivatives had to carry a notice that they were built using Falcon LLM technology from the Technology Innovation Institute.
On May 31, the commit log shows two changes in quick succession, one removing the TII Falcon LLM license and one updating the license info to Apache 2.0. A Hugging Face account updated the metadata tag the same day. The whole life of the royalty license was five days.
What the model was
Falcon-40B is a 40 billion parameter decoder trained on one trillion tokens, of which about three quarters is RefinedWeb English, a filtered and deduplicated extract of Common Crawl, with the rest split across European web text, books, conversations, code and technical sources. Training ran on 384 A100s for around two months from December 2022, which the Hugging Face post puts at 2,800 petaflop-days, roughly half the compute of LLaMA-65B. There is also a 7B model trained on 1.5 trillion tokens. Both use multi-query attention, which shrinks the key-value cache by a large factor and makes inference cheaper than a same-sized LLaMA.
On the Open LLM Leaderboard, which averages ARC, HellaSwag, MMLU and TruthfulQA, Falcon-40B-Instruct sits first at 63.2 and the base model second at 60.4. The Hugging Face post, written by Leandro von Werra, Lewis Tunstall, Pedro Cuenca, Philipp Schmid and colleagues, calls Falcon-40B the first truly open model with capabilities rivalling many closed ones. The phrase truly open is doing a lot of work in that sentence, and it is aimed squarely at LLaMA, which the same post's comparison table marks as not available for commercial use.
Why the clause mattered more than the score
Before the switch, Falcon was one more strong model that a company could not build a product on without a negotiation and a revenue share. That is the same category LLaMA occupies for a different reason, LLaMA being research-only, and the same category as every API model. The open community already had plenty of strong models it could look at but not ship. What it lacked was a good base it could fine-tune and sell without a lawyer, and Apache 2.0 gave it exactly that overnight.
You can see the effect in what Hugging Face did with the week. The blog post ships native support in transformers without trusting remote code, support in Text Generation Inference from version 0.8.2, and a QLoRA recipe that fine-tunes Falcon-7B on a single 16 gigabyte T4 with adapters of 65 megabytes. None of that engineering would have been worth doing for a model people could not deploy. The license change converted a leaderboard entry into a platform, and the ecosystem moved onto it within days.
That is the lesson we would draw about licenses as a lever. A model's rank on a leaderboard is a fact about the model. Whether a model becomes the default base everyone builds on is a fact about the model, the tooling, and the terms, and the terms are the cheapest of the three to change. TII spent months and thousands of GPU-days to reach second place. It spent one commit to become the model of the month.
What we would watch
The obvious question is whether this holds. Apache 2.0 is a real open source license with no takebacks, so the weights that exist now stay free regardless of what TII does next. But a lab that tried a royalty model once and abandoned it under pressure may try again with a different shape, and larger models are where the pressure to recoup training cost will be greatest. If a bigger Falcon arrives with terms that are permissive for small users and restrictive above a revenue threshold, that would be the royalty license coming back in a friendlier coat.
The other thing to watch is whether Apache 2.0 becomes the price of admission. As of this week, the top of the leaderboard is a commercially usable model and the model everyone was fine-tuning last month is not. If that pattern repeats, then any lab releasing weights under a custom license is choosing to be second in the ecosystem even when it is first in the benchmarks. We would like to see someone count, six months from now, how many derivative models on the Hub descend from Falcon versus LLaMA. That number would tell us whether the clause moved the leaderboard for a week or for good.
Sources
From the foundation