The Two-Word Instruction That Freed 402GB

"Remove old weights too"

At first glance, this message from the user in [msg 2091] appears trivial — a mere three words appended as an afterthought. But in the context of the opencode session, this brief instruction represents a critical inflection point: the moment when a project pivot transitions from intention to action, and when resource constraints force a hard取舍 between two massive models that cannot coexist on the same storage.

The Context: A Failed Experiment and a New Direction

To understand why this message was written, we must trace the arc of the preceding conversation. The session had been consumed by a multi-day effort to deploy GLM-5, a 744-billion-parameter Mixture-of-Experts model, on a machine with 8 NVIDIA RTX PRO 6000 Blackwell GPUs. The journey had been arduous: the assistant had patched vLLM's GGUF loader to support the novel glm_moe_dsa architecture, debugged Triton MLA sparse attention backends for the new SM120 compute capability, resolved tensor parallelism sharding mismatches that produced incoherent output, and ultimately achieved a working deployment with tool calling, 200k context, and ~57 tok/s throughput.

But the model's quality was poor. The user had chosen a 4-bit GGUF quantization (UD-Q4_K_XL), and in [msg 2088] they delivered the verdict: "the model was pretty unusable in that quant." The VM had been snapshotted to preserve the state, and a new experiment was to begin with nvidia/Kimi-K2.5-NVFP4 — a 1-trillion-parameter MoE model based on the DeepSeek V3 architecture, quantized by NVIDIA using NVFP4. The user's assessment was optimistic: "should be much simpler — try on latest vllm as that should run it natively... Also should be a bit faster because it's less active params and less layers."

The Assistant's Response and the Gap It Revealed

The assistant responded to this pivot in [msg 2089] by researching the new model. It fetched the HuggingFace model card for Kimi-K2.5-NVFP4 and, in [msg 2090], began surveying the current state of the machine. The results painted a clear picture of what needed to happen:

Why "Remove Old Weights Too"?

The user's message in [msg 2091] — "Remove old weights too" — is a corrective nudge. It reveals several layers of reasoning:

First, the user was monitoring the assistant's progress. They had read the assistant's research output and saw that the weight cleanup was listed as pending. Rather than wait for the assistant to eventually get to it, the user proactively ensured it would happen.

Second, the user understood the resource constraint. The new Kimi-K2.5-NVFP4 model is approximately 540 GB (as would later be confirmed when it was downloaded across 119 safetensor shards). With only 1.3 TB free on the shared storage, keeping a 402 GB dead weight would consume nearly a third of the available space — and the new model alone would take over 40% of the remaining capacity. The math is straightforward: 402 GB + 540 GB = 942 GB, leaving only ~358 GB for cache, temporary files, and headroom. The user recognized that the old weights were not just useless — they were actively obstructive.

Third, the user was thinking about operational hygiene. The phrase "too" is telling: it implies that removing the old weights is an additional step beyond what the assistant had already planned. The assistant had stopped the service and freed GPUs, but the user wanted the cleanup to be complete. Leaving orphaned weights on disk would be sloppy — they'd serve no purpose, consume space, and potentially cause confusion if someone later wondered why a 402 GB GGUF file was sitting in a directory named after a decommissioned model.

The Assumptions Embedded in This Message

The user made several assumptions, all of which were correct:

  1. The old weights were safe to delete. The VM had been snapshotted, meaning the entire state — including the GLM-5 weights — was preserved externally. Deleting them from the live environment was risk-free.
  2. The assistant knew where the weights were. The assistant's own research output had confirmed the path /shared/glm5-gguf/GLM-5-UD-Q4_K_XL.gguf and its size.
  3. Disk space would be needed. The new model was large, and while the exact size wasn't yet confirmed (the download hadn't started), the user correctly anticipated that 402 GB of headroom would be valuable.
  4. The assistant would act on this instruction immediately. The user didn't specify how to remove the weights — no rm command, no mv to a backup location. They trusted the assistant to execute the appropriate action.

The Outcome: From Instruction to Action

The assistant's response in [msg 2092] acknowledged the instruction and updated the todo list. In [msg 2093], the first bash command stopped the old service and checked the disk state. By [msg 2094], the GPUs were freed (0 MiB across all 8), and the old weights were confirmed present at 402 GB. The subsequent messages would show the assistant deleting the weights and proceeding to download the new model.

What This Message Reveals About the Collaboration

This brief exchange illuminates the dynamics of the human-AI partnership in this session. The user is not a passive observer — they are actively steering the process, catching details the assistant might defer, and ensuring resource constraints are respected. The assistant, for its part, had correctly identified the need but was operating sequentially: stop the service first, then remove weights. The user's intervention collapsed that sequence, ensuring cleanup happened before the new download began.

It also reveals the practical realities of working with frontier-scale models. A 402 GB file — larger than most entire datasets — is treated as disposable waste, deleted without ceremony to make room for the next experiment. The VM snapshot provides safety, but the casualness of the instruction ("Remove old weights too") speaks to the scale at which this environment operates. In a world where individual model weights consume half a terabyte, disk management is not an afterthought — it is a first-order constraint that shapes every decision.

Conclusion

The message "Remove old weights too" is a masterclass in concise, context-rich communication. Three words carry the weight of an entire project pivot: the failure of the GLM-5 quantization, the promise of the Kimi-K2.5-NVFP4 alternative, the arithmetic of disk space, the safety of snapshots, and the urgency of moving forward. It is a reminder that in complex technical collaborations, the most impactful instructions are often the shortest — because they rest on a foundation of shared understanding built over hours of preceding work.