Open-Weights Deployment

What you actually need to self-host the major open-weights models. VRAM per quantization (FP16, FP8, NVFP4, AWQ INT4, GGUF Q4_K_M), recommended GPU class, license, capabilities. This is the “I want to run this myself” side of the market. Rent it instead via /inference-providers (hosted open-weights pricing) or /models (frontier closed-model pricing per token). The closed bar these models are measured against moved on July 24, 2026: Claude Opus 5 shipped at $5/$25 per 1M tokens, holding Opus 4.8 pricing while posting 70.6 percent on OSWorld 2.0 and 90.8 on BrowseComp (vendor-reported), so the open-vs-closed capability gap widened at an unchanged closed-model price. It moved again on September 22, 2026, when Claude Opus 5.5 shipped at $4/$20 and took first place on the Artificial Analysis Intelligence Index at 58, where the top open-weights result as of late September, Xiaomi's MIT-licensed MiMo-V2.6-Pro, scored 46. October brought a wave of large open releases: Meta's Muse Spark Open (816B total, 42B active, under the custom Muse Model License with an access form), Reflection Beam (501B total, 23B active, Apache 2.0 with no gate), the Arabic-first Jais 3 family (Apache 2.0, no hosted API), and DeepSeek's V4.1-Flash-1017 weights update. Mistral Large 4 is not open weights yet: weights are promised by the end of October with no license named. Compare open models on capability at /best-open-source-llms.