Open-weight contenders: Llama, Qwen and the DeepSeek shock

Meta's cadence

Meta released Llama 3 on 18 April 2024, followed by Llama 3.1 on 23 July 2024, which the company described as its most capable models to date. Llama 3.2 arrived on 25 September 2024 and widened the family in two directions at once: 11B and 90B vision-capable models, and 1B and 3B lightweight models intended for on-device and edge deployment. On 5 April 2025 Meta released the Llama 4 herd, including Scout — a mixture-of-experts design with 17B active parameters — and Maverick, with a larger Behemoth model announced alongside them.

The edge-sized releases are the underappreciated part of that sequence. A 3B model that runs on a clinician's tablet without a network round trip solves a data-governance problem that no amount of contractual assurance solves for a hosted API.

“Open weights are not the same as open source, and procurement language should not be allowed to conflate them.”

DeepSeek and the cost question

Alibaba published the Qwen2.5 family on 22 November 2024, including specialised coder and mathematics variants, and launched Qwen3 with hybrid reasoning on 29 April 2025. DeepSeek published DeepSeek-V3 on 26 December 2024 and released DeepSeek-R1 on 20 January 2025 under an MIT license, claiming reasoning performance comparable to OpenAI's o1.

R1's release provoked a sharp market reaction, and the reason was economic rather than technical. A permissively licensed model with credible reasoning performance sets a ceiling on what closed providers can charge for the same capability class. Whether the reported training economics are reproducible is a separate question and remains contested; the pricing pressure was immediate regardless.

• 18 Apr 2024 — Llama 3; 23 July 2024 — Llama 3.1; 25 Sept 2024 — Llama 3.2 with vision and edge models.
• 5 Apr 2025 — Llama 4 Scout and Maverick released, Behemoth announced.
• 22 Nov 2024 — Qwen2.5 family; 29 Apr 2025 — Qwen3 with hybrid reasoning.
• 26 Dec 2024 — DeepSeek-V3 published; 20 Jan 2025 — DeepSeek-R1 released under MIT license.

Procurement guidance

Open weights are not the same as open source, and members should not let procurement language conflate them. Most releases in this category publish parameters under a bespoke license with use restrictions, without the training data or the full training recipe. That is enough to run, fine-tune and audit behaviour; it is not enough to reproduce the model or to fully characterise what it learned.

For regulated deployments, the practical checklist is: can we run it inside our trust boundary, can we pin a version indefinitely, can we evaluate it against our own held-out data, and can we accept the license terms for our commercial use? Where all four are yes, an open-weight model is often the lower-risk choice even when a hosted model scores higher on public benchmarks.