One Model to Rule Them All: Qwen3.5 122B on a Homelab for CF Agents | CF Weekly #85 скачать в хорошем качестве

One Model to Rule Them All: Qwen3.5 122B on a Homelab for CF Agents | CF Weekly #85

You don't need the cloud. You need four RTX 3090s and a dream. This week on CF Weekly we deploy Qwen3.5-122B-A10B — 122 billion parameters, 10B active per token, 262K native context — directly on homelab hardware using vLLM. We cover the GPTQ Int4 quantization gotchas, the `moe_wna16` flag nobody tells you about, and how this MoE beast slots into agentic CF workloads. But we don't stop there. We put it to work with OpenCode, wire it into a BOSH-deployed OpenClaw instance so everyone can have their own AI agent platform on Cloud Foundry, and — if the demo gods are with us — throw it into a head-to-head cage match against GPT-OSS-120B to see which model truly earns the crown. One model. One rig. One Ring. No cloud bill.