The open-weights wave: months, not years, behind
The growing availability of large language model (LLM) weights has shifted the balance of power in AI development. This lesson explores the current state of open-weight models and their performance relative to closed systems. It identifies the main open-weight model families and quantifies the gap between open and closed models using empirical metrics.
Open-weight model evolution
The inflection point in open-weight development came with DeepSeek R1 in January 2025. This 671-billion-parameter mixture-of-experts (MoE) model activated 37 billion parameters per token. It matched o1-class reasoning on mathematics and coding benchmarks while requiring a fraction of the training cost. The model’s weights were fully downloadable, setting a new standard for accessibility. Subsequent models such as V3.2 (late 2025) and V4 (April 2026) further advanced the field. V4 introduced a one-million-token context window and agent-oriented post-training capabilities.
China’s open-weight contributions
China’s open-weight models have made significant inroads, particularly with Zhipu’s GLM-5 series. Released in February 2026, GLM-5 totalled 744 billion parameters with approximately 40 billion active parameters. It was MIT licensed and trained on 100,000 Huawei Ascend chips without any US hardware. GLM-5 achieved a 77.8% SWE-bench Verified score, posting the strongest open coding results of its time. Successive versions, GLM-5.1 and 5.2, continued this trend. Moonshot’s Kimi K2 line also reached a trillion total parameters, with 32 billion active parameters and an 80.2% SWE-bench Verified score. The Kimi K2 line was released under the Apache 2.0 license.
Western open-weight developments
The Western open-weight landscape was dominated by Meta’s Llama 4 family, which introduced MoE capabilities to the Llama series. Scout, part of this family, featured 109 billion active parameters with 17 billion active parameters and a 10-million-token context window. Maverick, another Llama 4 model, featured 128 experts. Meanwhile, Mistral maintained its Apache 2.0 flagships. The company’s revenue grew from 20 million to 400 million dollars annual recurring revenue (ARR) within a year, demonstrating the commercial viability of open-weight models.
Open-weight performance and licensing
Qwen 3.x models completed the open-weight landscape with Apache 2.0 licensing. These models demonstrated reasoning capabilities competitive with closed-flagship models in benchmark suites. The licensing terms allowed for broad use and integration. The measured performance gap between open and closed models was quantified by Epoch AI in 2026. Their analysis placed the open-weight frontier 4 months behind the closed frontier on its capability index. Under stricter criteria, the gap was 6 months, about 8 ECI (Effective Capabilities Index) points. This gap represented a slight widening from earlier 3-month estimates but remained historically small.
Strategic implications
The strategic consequence of this development is significant. When model weights are nearly free and nearly frontier, the competitive advantage shifts from model development to inference stack capabilities. Entities that can run these models well, cheaply, and under their own control gain a distinct edge. This shift emphasizes the importance of efficient deployment and inference infrastructure. It also highlights the growing importance of custom inference stacks and optimization techniques.
What to take away
Open-weight models have reached a critical stage where their capabilities closely match closed systems. The measured performance gap is now around 4 months, with some metrics showing a 6-month lag under stricter conditions. The advantage has shifted from model weights to efficient deployment and inference capabilities. This transition signals a new phase in the AI landscape where infrastructure and optimization matter as much as model architecture.
Reference
| Lesson | 3 of 15 |
| Outcome | Name the open frontier model families and quantify the open-closed gap. |
| Worked engine | Colibri disk-streaming MoE engine |
| Gap tracking | Epoch AI open-vs-closed analysis |