Run an open model as a long agent (a model that calls tools in a loop on its own) on one card and the weights are rarely what stops you. The budget around them is: how much the agent can keep in view, and how long one answer may be.
We learned that with SenseNova 6.8 Code2Video, SenseTime's open 27B model that makes video by writing code. We drove it with Claude Code on one 24 GB RTX 4090 on our cloud. Ten minutes in, it had a full 12-second clip. Forty minutes in, it had no finished film. The logs and frames are in the full research write-up on running Code2Video on one RTX 4090. Here I focus on the settings.
Your settings are not the authors' settings
The SenseNova team publishes a Claude Code skill called code-video and six complete production runs. They serve the full-size model across four GPUs. Their context window (the text the model can see at once, counted in tokens, which are word pieces) is 512K tokens, and one answer may run to 64K. Their films needed between 261 and 876 turns each.
We used bartowski's 4-bit GGUF, a compressed copy of the weights (17.2 GB) for the llama.cpp server, plus a 0.93 GB vision part so the model can see images. Three Claude Code variables encode the gap:
CLAUDE_CODE_MAX_CONTEXT_TOKENS=98304 # authors: 524288 CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80 # same as the authors CLAUDE_CODE_MAX_OUTPUT_TOKENS=16384 # authors: 64000 Without the first one, Claude Code assumes 200K for a model it does not recognize.
The trap: compaction needs room to write
Compaction is how Claude Code survives a full context. Once the window passes 80%, it has the model write a recap, throws the old history away and continues from the recap alone. The recap is a normal reply, so the output cap applies to it too.
Our 98,304-token window was full after about 19 minutes. Three recap requests followed. All three ran into the 16,384-token ceiling, none produced a usable summary, and together they ate 21 minutes until our 40-minute limit killed the run with no final film.
Compaction can work: four of the authors' six runs cleared contexts of 167K to 492K tokens, under a 64K cap. On our side two settings are under suspicion. One is the 16,384-token cap. The other is hidden reasoning, the thinking the model writes before its visible answer, which we left on and which draws from the same cap. The logs cannot tell them apart, though in a separate single request the same 4-bit file burned all 8,192 output tokens thinking and wrote no page. A bigger window, a bigger cap or reasoning off might fix it; we tried none of them.
So when you size a single-card agent, measure two things before you trust it: when the context will hit the compaction threshold, and whether one summary fits under your output cap with reasoning on.
The file check passed a wrong picture
Self-review did its job. The model opened 21 of its own stills as images, spotted two bugs and fixed both. All on-screen text matched the brief word for word.
The one fact that mattered was still wrong. The brief asked for 30 dots in a 5 by 6 rectangle. We located each dot centre in the 7.2-second frame with numpy and scipy and got 5 rows of 7 columns, one cell empty per row. The skill's file check still said yes. It only looks at resolution, frame rate and duration, never at what is drawn.
Add a check that knows what the clip should show.
Two pieces of plumbing to verify first
Claude Code speaks Anthropic's Messages API format, so whatever sits in front of the model has to translate it. Ours lost the frames: each image either vanished or arrived as a block of text, and the model reviewed its work blind. Fixed in our test build, not yet on the hosted service. With any gateway, send one PNG first and check the model describes it.
Second, time. Our first call waited for the full answer instead of streaming it piece by piece. The worker, the process on the GPU machine that forwards requests to the server, gives up after 120 seconds, and the model was still thinking. Raising that to 30 minutes fixed the run, and the example now streams.
What fits on smaller cards
The 4-bit run peaked at 20.9 GiB. Smaller copies, one short request each:
| Budget | Copy | Peak | Result |
|---|---|---|---|
| 16 GB | 3-bit, 32K | 13.7 GiB | page renders, one dot visible at a time |
| 12 GB | 2-bit, 16K | 10.0 GiB | page cut off at 8,192 tokens |
These ran on the 4090, not real 16 or 12 GB cards, with one sample each. The memory ladder, launch command and step-by-step setup for the Code2Video test have the rest. See also our earlier test of another SenseTime model, SenseNova-Loop-9B.
Credits and the honest limit
The SenseNova team at SenseTime built the model, the skill, the prompt template and the production runs, starting from the Qwen team's base model (model card, skill repository). bartowski made the GGUF quants and the llama.cpp maintainers made the server. Anthropic made Claude Code and Claude Opus 5.5, which drafted the brief from the authors' template. We first saw the model in an AI Search news video. We handled the deployment, fixed the image path, ran the sandboxed agent and counted the pixels.
This was one attempt at one brief under tight limits we chose. Logs, timings, the full VRAM table and the SDK example are in the complete Code2Video research post on inferencekey.com.

