The open-weight versus closed AI model gap

Mozilla's State of Open Source AI report, cited by Ars Technica, puts the performance gap between closed frontier models and the best open-weight models at 4.4 months. Moonshot AI's Kimi K3 scores three points below Anthropic's Fable 5 on the Artificial Analysis Intelligence Index at roughly 30% of the cost. Mozilla's recommendation is that the decision to pay for closed models should be workload-specific rather than organization-wide, reserving the closed model premium for expert professional work, high-intensity retrieval, and long-context tasks. Independent analyses from SemiAnalysis, Epoch AI, and Artificial Analysis estimate the gap at roughly 4 to 6 months, with leading open models coming predominantly from Chinese labs since approximately 2024.

Nvidia CEO Jensen Huang disputes the cost framing, arguing closed models are actually cheaper for most users because self-hosting open models requires significant cost and expertise to train, fine-tune, maintain, guardrail, and evaluate them. Huang said the primary reason to use open models is control, not cost savings.

Recent open-weight releases include: IFM's K2 Horizon (six models spanning 0.9B to 375B parameters, with full release of weights, training code, data recipes, checkpoints, and the complete agentic post-training codebase); Zhipu AI's GLM-5.3-Flash (ranking in the top 5 of 112 large open-weight models on the Artificial Analysis Intelligence Index with 18B active parameters and a 1 million token context window); Cohere's North Small Translate (an open-weights 218B/25B-active mixture-of-experts model scoring 83.5 on WMT26, above DeepL NextGen's 81.2); Qwen's Qwen3.8-Flash-Next (an open-weight multimodal mixture-of-experts model) alongside an early preview of Qwen4 described as capable of running locally on a single NVIDIA DGX Spark; Z.ai's GLM 5.3 (a 744-billion-parameter model tuned for multi-hour agent sessions on DGX Station and four-node DGX Spark clusters); and Tencent's Hy4-preview (larger than its predecessor Hy3 with an improved post-training process, but currently carrying a known overthinking issue).

US lawmakers have probed DoorDash, Airbnb, Anysphere/Cursor, and Apple over their use of Chinese AI models. Perplexity prominently adopted DeepSeek R1, and Thomson Reuters built on Qwen to move off Claude, citing cost savings.