Open source models close the gap: DeepSeek V4, Qwen 3.7 Max, Llama 4
Open-weight models now trail frontier proprietary systems by single-digit percentages on everyday tasks, while costing four to ten times less per token.
NFNiko Ferrar
Published 6 Jul 2026, 11:34Updated 28 Jul 2026, 19:111 min read
A wave of 2026 open-weight releases has narrowed the practical gap with closed frontier models to single-digit percentage points on everyday tasks, according to comparative benchmark roundups.
The models leading the pack
- ▸DeepSeek V4, released in March, is described as the model that made enterprise architects take open source seriously for reasoning-heavy, long-context workloads
- ▸Qwen 3.7 Max has become the broadest open-source family by capability and the most-downloaded open model line to date
- ▸Llama 4 Scout and Maverick ship as open-weight multimodal models under Meta's Llama 4 Community License
Why the cost gap matters more than the capability gap
Credit costs for these models run four to ten times cheaper than premium proprietary tiers, according to the roundups. For teams whose workloads sit inside that single-digit capability gap, the economics increasingly favor open weights over frontier APIs.
Source: Hugging Face
