Open-weight models at near-frontier quality

Several open-weight models now reach near-frontier benchmark performance while costing far less than closed alternatives. Kimi K3, a 2.8 trillion parameter mixture-of-experts model from Moonshot AI released July 27, 2026, scored 88.3 on Terminal-Bench 2.1, placing it close to GPT-5.6 Sol (88.8) and above Claude Fable 5 (84.6), according to Cline.bot. Fastino.ai reports Kimi K3 reaches 95% of Fable 5's intelligence score at 35% of the cost, while GLM-5.2 from Z.ai delivers 85% of that score at 17% of the cost. V4 Pro scores 80.6% on SWE-bench Verified, placing it in the same range as Claude Opus 4.6 and Gemini 3.1 Pro for coding tasks. DeepSeek's updated model scored 82.7 on Terminal-Bench 2.1, a 25.8-point improvement over its April 2026 preview score of 56.9, and added support for tool calls and the Responses API. OpenRouter reports open-weight models have maintained a consistent 3-6 month capability gap behind frontier labs for over 18 months, and that cost pressures are driving enterprise adoption. NVIDIA's Nemotron 3 Ultra is identified by OpenRouter as the strongest US-built open-weight model, described as a serious reasoning model for enterprise deployment backed by NVIDIA's hardware and software stack. MindStudio notes inference providers like Together AI, Fireworks, and Groq offer open-weight access at typically 10-15x less than frontier APIs. Fastino.ai argues that open weights make published API prices a ceiling rather than a floor, since any provider can serve these models and drive costs lower through competition. Closed frontier models retain an edge on complex reasoning, multimodal tasks, reliable tool use, and fast prototyping, according to MindStudio. On the local deployment side, Microsoft showcased DeepSeek V4 Flash, a 284-billion-parameter model, running on a Windows PC after quantization to 1.6 bits reduces its memory requirement to approximately 60GB; the Surface Laptop Ultra with up to 128GB of unified memory on Nvidia's RTX Spark chip can accommodate it.