A single 24GB GPU can now run a 35B scientific research agent that spawns sub-agents and checks its own evidence. Sage-A1.5 is 35B / ~3B active, and its official Q4_K_M is just 21.7GB. Here are the new Sage-A1.5 model stats 🧠 35B MoE / ~3B active 💾 official Q4_K_M → 21.7GB 🎮 fits a single 24GB GPU at moderate context 📚 256K native context 👁️ multimodal 🔓 Apache 2.0 And it's specifically trained as a long-horizon research agent. So it is designed to 🔎 use tools/search/code 🤖 spawn parallel sub-agents ✅ inspect its own evidence 🔄 change direction when research stalls No separate critic model required for its verification mode. These are the full-precision benchmark results PinchBench Sage-A1.5 → 90.0 Kimi K3 → 88.2 GLM-5.3 → 86.6 GPT-5.6 Sol → 87.3 FrontierScience-Olympiad Sage → 81.0 Kimi → 69.0 GLM → 67.0 GPT → 73.0 SWE-Pro Sage → 61.2 Kimi → 63.3 GLM → 64.6 GPT → 64.6 ⚠️ Those benchmark scores are from the BF16 checkpoint, not the Q4. But the same model now has an official 21.7GB GGUF with llama.cpp / Ollama / LM Studio support. Link in ALT