LM Studio, Codex integration hits VRAM wall

The reported "stopping" is likely a connection timeout triggered by the Codex app's agent harness when token generation speeds drop below a hardcoded threshold during VRAM-to-RAM paging.

• –Qwen 3.5 27B at standard 4-bit quantization requires ~20GB of VRAM; even with an RTX 5070Ti and ADA 2000, unoptimized layer offloading in LM Studio can cause severe latency spikes.
• –Windows Hardware-Accelerated GPU Scheduling (HAGS) frequently conflicts with dual-GPU LLM setups, leading to the exact "midway stop" behavior described.
• –The Codex desktop app (released early 2026) uses a bidirectional JSON-RPC protocol that is less tolerant of the high-latency "chugging" common when models spill into system DDR5 memory.
• –Switching to the Qwen 2.5 Coder 14B model is the recommended "sweet spot" for this hardware, ensuring the entire KV cache remains on-chip for stable agentic workflows.