Skip to content
NetProjects
Menu

Insights

SenseNova 6.8 Code2Video on a Mac: a 27B model that writes videos as code

NetProjects

Most video models generate pixels. SenseNova 6.8 Code2Video writes the film as code instead: it plans a piece, writes the scenes, renders them, looks at the frames and revises, turn after turn. The result is a short video you can inspect, edit and rerun, because every frame comes from a program.

It is a 27B agentic model, fine-tuned from Qwen3.8-27B and released with open weights under the Apache-2.0 licence. We ported it to MLX so it runs on Macs with Apple Silicon, with text and images, and we checked every build against the full-precision weights. The builds are on Hugging Face.

Always up to date. We keep these builds current on Hugging Face. When SenseNova releases a new version of SenseNova 6.8 Code2Video, we rebuild, re-test and publish it there, so the links on this page always point to the latest version. The numbers in this article were measured on the version named in each section; the model cards on Hugging Face carry the current figures.

What the model does

SenseNova describes the model as making short videos entirely in code: motion design, explainers, kinetic type and data stories. It can also call other tools through code, such as image or video generators, when a scene needs them. The rendering tools, craft guides and prompts are in SenseNova's code-video skill: https://github.com/OpenSenseNova/SenseNova-6.8-Code2Video. SenseNova's original is at https://huggingface.co/sensenova/SenseNova-6.8-Code2Video.

SenseNova also released production trajectories from real runs, which we used in our checks. It says a complete film can run for many hours and thousands of turns; we have not timed a full production run ourselves.

What we released

Four MLX builds on Hugging Face, converted from SenseNova's weights at revision 0f99ab0. The quantised builds quantise the language model, affine with group 64, and keep the vision tower in bf16.

Hugging FaceSize
netprojects/SenseNova-6.8-Code2Video-MLX-6bit (recommended)22.8 GB
netprojects/SenseNova-6.8-Code2Video-MLX-8bit29.5 GB
netprojects/SenseNova-6.8-Code2Video-MLX-4bit16.1 GB
netprojects/SenseNova-6.8-Code2Video-MLX-bf1654.7 GB

We tested the builds with mlx 0.32.4, mlx-lm 0.32.2 and mlx-vlm 0.7.11. The YaRN rope scaling, factor 2, from the original config is kept, so long contexts work as in the original model. The original checkpoint ships without a multi-token prediction head, so these builds have none either.

How well it matches the original

We first checked that the weight files we downloaded from SenseNova match the hashes published on Hugging Face. Then we compared each quantised build with the bf16 MLX build, the full-precision weights, on 20 assistant turns from SenseNova's released production trajectories, with the chat template, tool schemas and up to 16,000 tokens of context each, 10,406 scored tokens.

Measured on SenseNova 6.8 Code2Video, 16 November 2026, MacBook Pro M5 Max (128 GB).

BuildSame next token as bf16, agent turnsMean NLL vs bf16, agent turnsSame next token as bf16, 12 prompts
MLX bf16ReferenceReferenceReference
MLX 8-bit98.3%−0.00199.1%
MLX 6-bit96.8%±0.00098.7%
MLX 4-bit91.7%+0.03297.3%

The 12-prompt figures are teacher-forced over 878 positions. NLL measures how far a build's token probabilities drift from bf16; closer to zero is closer to the original. On real agent turns the 6-bit and 8-bit builds stay close to bf16, which is why we recommend them. The 4-bit build is the smallest option; it stays coherent but drifts more from bf16 on long agent turns.

On the 8-bit build, a fact placed in a 65,586-token prompt is retrieved exactly. On a request with a tool schema, the 8-bit build returns one well-formed tool call with the right function and arguments, and the 6-bit build, served through mlx_vlm.server, returns the same request as a parsed OpenAI tool call. The 4-bit, 6-bit and 8-bit builds each describe a test image correctly in mlx-vlm.

Try it

On a Mac with Apple Silicon, with mlx-vlm, for text or text with an image:

pip install -U "mlx-vlm>=0.7.11"
mlx_vlm.generate --model netprojects/SenseNova-6.8-Code2Video-MLX-6bit --max-tokens 512 \
  --prompt "Plan a 10-second kinetic-typography intro for the word ORBIT: scenes, timing and colours."

mlx_vlm.generate --model netprojects/SenseNova-6.8-Code2Video-MLX-6bit --max-tokens 256 \
  --image frame.png --prompt "Review this rendered frame: what would you change in the layout?"

mlx_vlm.server serves an OpenAI-compatible chat API with tool calls:

mlx_vlm.server --model netprojects/SenseNova-6.8-Code2Video-MLX-6bit --port 8100

Point your agent harness at the local endpoint and give it SenseNova's code-video skill. SenseNova's recommended harness settings for long runs are on the original model card: https://huggingface.co/sensenova/SenseNova-6.8-Code2Video. Run the code the model writes in a sandbox you control, and review renders before you publish them.

Honest limits

  • All figures come from one machine.
  • We have no speed figures, and no memory figures beyond the file sizes above.
  • We have not rendered a complete video or run SenseNova's full code-video skill end to end.
  • We checked long context only on the 8-bit build, with a 65,586-token prompt, not the 512,000-token context SenseNova describes.
  • We checked images only by asking each quantised build to describe one test image.
  • The 4-bit build differs more from the original than the others.

Licence

SenseNova 6.8 Code2Video, Qwen3.8-27B and our builds are all under the Apache-2.0 licence. Each of our repositories carries the Apache-2.0 text and a notice that describes the conversion and credits SenseTime SenseNova and Alibaba Cloud.

Links

Always up to date. We keep these builds current on Hugging Face. When SenseNova releases a new version of SenseNova 6.8 Code2Video, we rebuild, re-test and publish it there, so the links on this page always point to the latest version. The numbers in this article were measured on the version named in each section; the model cards on Hugging Face carry the current figures.

Credits

SenseNova 6.8 Code2Video is by SenseNova, fine-tuned from Qwen3.8-27B by Alibaba Cloud's Qwen team, under the Apache-2.0 licence. The MLX builds are by NetProjects (Internet Projects Ltd). Not affiliated with SenseTime or Alibaba Cloud.

This article was drafted with AI assistance, then edited and approved by people on our team. You can read more about how we use AI at https://netprojects.com/standards/ai.