Recommendation List for Trending Open Source Model Providers (docs.google.com)

This recommendation ranks trending open‑source model providers by evaluation metrics, time‑to‑first‑token (“aha” moment), pricing, uptime and quantization, and is intended to drive Cline-provider’s Open Router routing decisions. Standouts: Groq (no quantization) delivers the best perceived quality and the fastest experience (≈1058 tps, 0.2 s latency) and uniquely supports prompt caching — cutting effective cost from $1 input to $0.50 on cache hits and producing the “magical AHA.” Fireworks is a reasonable fp8 fallback (≈105 tps, 0.8 s, $0.60). For coding, the Qwen3 Coder (480B A35B) is recommended, with Nebius (fp8, 84 tps, 0.48 s, $0.40) and BaseTen (fp8, 104 tps, 0.84 s, $0.38) as provider options—BaseTen is cheaper but has higher first‑token latency. Other notes: GLM 4.6 and variants are fp8/bf16 mixes with moderate throughput and varied latencies and prices (cache hits can drastically reduce per‑call costs). DeepSeek’s best model (deepseek-v3.2-exp) is slower (≈25–28 tps, 1.05–1.55 s) but very cheap post‑cache and enjoys heavy real‑world traffic. GPT-oss-120B shows both very high throughput configurations (≈954 tps, 0.27 s, $0.15; cache can drop cost further) and slower, ultra‑cheap options. Bottom line: route by workload — prioritize no‑quant / low‑latency providers and prompt caching for latency‑sensitive, high‑quality use cases; use fp8/quantized providers for cost‑sensitive or fallback capacity.