← Back to blog
AiAbout 7 min read

Making video out of code just got picked up by a 27B open model

Published Nov 7, 2026
Making video out of code just got picked up by a 27B open model

There's a post on Xiaohongshu that got saved 1,233 times in early November, with 47 people going at each other in the comments. The claim is blunt: "Opus took 291 yuan to make one 3D motion piece. I switched to SenseTime's open-source Code2Video and it cost 2.8. A hundred times cheaper."

Posts like that usually don't last three days, because they're too easy to argue with. The top comment does the job nicely: once you count debugging and rework, staying on Opus 5.5 might well be cheaper. The low unit price is bought with time and iteration count.

But there's something behind that post worth a real look. On October 28, SenseTime put the weights for SenseNova 6.8 Code2Video on Hugging Face: 27B parameters, 512K context, Apache-2.0 on the model card. What it does isn't the same category as Sora or Runway at all. It doesn't guess pixels. It writes code.

Every frame is a function of time

A conventional video model samples the whole clip in one pass. You want the subtitles two points bigger, so you re-prompt and re-sample, and the worst part is that the things you didn't ask to change may change along with them. The face stops matching.

Code2Video comes in through a different door. It writes HTML/CSS, Canvas, SVG, Three.js, Manim or p5.js, and defines each frame as a pure function of time t: give it any point in time and the program computes, uniquely, the position, deformation and colour of every element at that millisecond. Rendering captures frame by frame, with no dependence on the system's real-time clock. Run the same code a hundred times and the output is identical.

A loop of blank 35mm film frames glowing faintly on a lit glass table, no markings.

The approach comes with a side effect: text and numbers stop smearing. Subtitles, formulas, the values inside a chart are typeset from source code, not guessed as pixels. Anyone who's made AI video knows text in the frame is where things break first, Chinese especially.

Changing things is cheaper too. Elements in the code have names; change a font size, a timestamp, a palette, and you touch those few lines and nothing else. The official example is an airport split-flap board, revised three times: first a dark board under a spotlight, then a bright pastel concourse, then a plane and travellers dragging luggage. Across all three rounds, the copy, timing and flip motion nobody asked to change stayed exactly as it was.

SenseTime also published the prompts used to make the films and the full agent logs, pitfalls included. The repo ships a skills/code-video directory with 8 domain guides. Read end to end, it looks more like a textbook on teaching your own agent to make video than a model card.

Still 483 points short on the leaderboard

The value of this approach needs a leaderboard underneath it.

HeyGen updated a Kaggle board called Code2Video: give the model a detailed creative brief, have it write code inside the HyperFrames framework to produce a clip, then score it against a human-made reference. The top eight are all closed models. Claude Opus 5.5 is first at 33.5% raw. Open models like GLM, Kimi and DeepSeek sit around 10%.

Artificial Analysis reproduced the evaluation with a finer measure: Opus 5.5 scores 1536 Elo, the strongest open general model DeepSeek V4.1 Flash only 1053. A gap of 483.

One line in the leaderboard notes I agree with: plenty of frontier models can write HTML, CSS and JS, but turning code into a reliable video means understanding the frame along the time axis, not just drawing a pretty static page. Writing a web page isn't shooting a film. A static page wins by placing elements correctly; a video has to make elements move at the right moment, at the right speed, in the right order — when the title enters, how the number blocks jump with the beat, which frame the drum hit lands on.

So the hollow middle was awkward: want quality, pay for a closed model at tens of yuan a clip with no local deployment; want open source to save money, and general models weren't trained for video. Code2Video fills that space, by specialising.

Three bills to run first

The first is hardware. 27B parameters, BF16 weights around 54GB. The official docs give three deployment commands, LightLLM, vLLM and SGLang, all with tensor parallel 4, four cards minimum, and 512K context needs extra KV memory. Sounds frightening, but the Qwen3.8-27B base is frugal: quantised to Q4 it's around 18GB, and a single 24GB 4090 will run it. On Douyin, a creator on a 64G Mac Studio used one prompt to have the AI detect the config, install the environment and deploy the model, then hooked it into Claude Code through CC Switch, barely typing a command by hand. Another ran the 4-bit quantised version on an ordinary machine with 32G of RAM.

The second is time. Officially it can run 10 hours plus, 2,000-odd turns. Read backwards, that sentence means: the job takes that long anyway. In practice a single finished clip takes tens of minutes to one or two hours, depending on local compute. This isn't one-click video. It's leave-it-running video.

The third is capability. The official limits say it plainly: motion graphics, explainers, stylised 2D/3D scenes, data visualisation, yes; photoreal footage, no. The Douyin creator tried a few things on it: fish schooling, turning in sync with no leader; drawing a cat with Fourier series, more circles and a sharper outline; prism dispersion; two-source interference; a chaotic double pendulum; a crank and connecting rod; a Möbius strip. Smooth, and the point gets across. All of them happen to be pictures you can write as code. If you want live-action texture and cinematic camera work, this isn't the tool.

The official docs even suggest handing prompt enhancement to a stronger model like Claude. That kind of honesty is rare.

The licence has two labels

"Commercial use is fine" deserves a question mark. The Hugging Face model card says Apache-2.0, but the GitHub repo root has no LICENSE file right now, and the GitHub API can't read a licence. The repo also carries the property of executing shell commands, and the README explicitly warns you to run it in an isolated environment.

As of November 4, the GitHub repo has about 390 stars and the HF model page 93 likes. In the trade's language, still very early. There's no systematic benchmark from the team either, only qualitative samples, and finish and long-range stability clearly trail the strongest closed models. What you can see is sample films, not scores.

If you're actually going to use it, one detail is worth copying: it ships a set of check scripts whose idea is to cold-read keyframes, generate contact sheets, and verify resolution, frame count, black frames and frozen frames. Acceptance only writes down what's visible in the pixels, not what's written in the code. That principle outlasts the model, because code that runs doesn't mean the picture is right — a formula off by one, subtitles cropped, the subject too small, and the compiler won't say a word.

What actually got cracked open is marginal cost

Back to that post. It isn't saying open source has caught up with closed. The leaderboard is right there, 483 points short, and that gap won't close soon.

What got cracked is the marginal cost. Someone ran Opus 5.5 across 9 parallel agents for 6 hours to make one trailer, burned 187 million tokens, and got a 190-dollar bill. Someone else spent 270 dollars on an 8.5-minute video, 200 of it rework from going back and forth. With closed models the expensive part isn't generation, it's rework: six rounds, six bills.

Once the model sits on your own machine, a hundred clips cost about what one costs; what's left is electricity. A difference of that size lands hard for three kinds of people: matrix accounts at the launch stage that need low cost and post explainers, data, rankings, where volume makes closed-model token burn frightening; state-owned and government enterprises that must deploy on the internal network, model, knowledge base and application all physically isolated, data not allowed to leave; and corporate trainers and teachers rebuilding courseware, where a 30-second animation beats a thousand sentences.

The turning point isn't who's stronger. It's who can do it over and over at zero cost. That happens to be the closed models' weak spot, and the place where a small model like 27B has the most to gain.

Related articles