SenseNova 6.8 Code2Video for Apple Silicon (MLX, 4-bit)

An MLX conversion of sensenova/SenseNova-6.8-Code2Video (revision 0f99ab0), SenseNova's 27B agentic model for code-driven video creation, fine-tuned from Qwen3.8-27B. It makes short videos by writing code: it plans a piece, writes the scenes as code, renders them with the rendering tools of SenseNova's code-video skill, looks at the frames and revises, turn after turn. The skill, craft guides and prompts are in SenseNova's GitHub repository. This build quantises the language model to 4-bit (affine, group 64) and keeps the vision tower in bf16. Size: 16.1 GB.

Published by NetProjects (Internet Projects Ltd). Not affiliated with SenseTime or Alibaba Cloud.

Also from SenseNova on MLX: SenseNova-Loop-9B-Preview MLX.

Build Size Same next token as bf16 (agent turns) NLL vs bf16 (agent turns)
netprojects/SenseNova-6.8-Code2Video-MLX-6bit (recommended) 22.8 GB 96.8% ±0.000
netprojects/SenseNova-6.8-Code2Video-MLX-8bit 29.5 GB 98.3% −0.001
netprojects/SenseNova-6.8-Code2Video-MLX-4bit 16.1 GB 91.7% +0.032
netprojects/SenseNova-6.8-Code2Video-MLX-bf16 54.7 GB reference reference

For agent work, use the 6-bit (recommended) or 8-bit build: on real Code2Video agent turns both score the same as bf16. The 4-bit build is the smallest option; it stays coherent but drifts more from bf16 on long agent turns.

Run it

pip install -U "mlx-vlm>=0.7.11"

Text, or text with an image:

mlx_vlm.generate --model netprojects/SenseNova-6.8-Code2Video-MLX-4bit --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-4bit --max-tokens 256 \
  --image frame.png --prompt "Review this rendered frame: what would you change in the layout?"

As a local endpoint for agent tools:

mlx_vlm.server --model netprojects/SenseNova-6.8-Code2Video-MLX-4bit --port 8100

mlx_vlm.server serves an OpenAI-compatible chat API with tool calls. The model writes Qwen3 <tool_call> blocks; 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 in the original model card.

The YaRN rope scaling (factor 2) from the original config is kept, so long contexts work as in the original model; we checked 64K tokens on the 8-bit build. The upstream checkpoint ships without a multi-token-prediction (MTP) head, so these builds have none either.

What was checked

On a MacBook Pro M5 Max (128 GB), 16 November 2026, with mlx 0.32.4, mlx-lm 0.32.2 and mlx-vlm 0.7.11.

  • Parity with bf16 on real agent turns. 20 assistant turns from SenseNova's released production trajectories (trajectories/all.openai.jsonl), with the chat template, tool schemas and up to 16K tokens of context each, 10,406 scored tokens:

    Build Mean NLL vs bf16 Same next token as bf16
    8-bit −0.001 98.3%
    6-bit ±0.000 96.8%
    4-bit +0.032 91.7%
  • Token parity with bf16. On 12 fixed prompts (878 teacher-forced positions), the builds pick the same next token as bf16 on 99.1% (8-bit), 98.7% (6-bit) and 97.3% (4-bit) of positions.

  • Long context. On the 8-bit build, a fact placed in a 65,586-token prompt is retrieved exactly.

  • Tool calls. On the 8-bit build, a request with a tool schema returns one well-formed <tool_call> with the right function and arguments (tested through mlx-vlm generate). On the 6-bit build, mlx_vlm.server returns the same request as a parsed OpenAI tool call (finish_reason: tool_calls).

  • Images. The 4-bit, 6-bit and 8-bit builds each describe a test image correctly in mlx-vlm.

  • Source files. All weight files of the upstream revision match the Hub's SHA-256 hashes.

Intended use

SenseNova 6.8 Code2Video is built for programmatic video: motion design, explainers, kinetic type and data stories made entirely in code. Run the code it writes in a sandbox you control, and review renders before you publish them. See the original model card for the model's scope, example films, the production trajectories and the recommended harness.

License

Apache-2.0, as the original model and Qwen3.8-27B. LICENSE is the Apache-2.0 text and NOTICE describes the conversion and credits SenseTime SenseNova and Alibaba Cloud.

More MLX builds from NetProjects: huggingface.co/netprojects · netprojects.com

Downloads last month
-
Safetensors
Model size
27B params
Tensor type
U32
·
BF16
·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for netprojects/SenseNova-6.8-Code2Video-MLX-4bit

Quantized
(9)
this model

Collection including netprojects/SenseNova-6.8-Code2Video-MLX-4bit