Quantization
This repository is an EXL3 6.0bpw quantization of sensenova/SenseNova-6.8-Code2Video.
Table with columns: Setting, Value| Setting | Value |
|---|
| Format | EXL3 (exllamav3 v1.6.5) |
| Bits per weight | 6.0 (head_bits: 6, vision_bits: 6) |
| Codebook | mul1 |
| Out scales | auto |
| Calibration | 250 rows x 2048 columns |
No MTP layer. The original model's config declares one MTP layer (mtp_num_hidden_layers: 1), but no MTP weights are present in this checkpoint. config.json was therefore modified from the original: text_config.mtp_num_hidden_layers was changed from 1 to 0, and text_config.mtp_use_dedicated_embeddings was removed.
Quantization tooling. The quant was produced with a modified copy of exllamav3's exl3_lib/quantize.py, included here as scripts/quantize.py (MIT-licensed exllamav3 code). Changes relative to the stock version: support for the packed block-packed-upper-triangle Hessian capture (decoded via hessian_packed.unpack), chunked and in-place Q/H statistics that drop the full k x n float32 intermediate (chunk default 16384 to 1024), Hessian factorization on the quantizing device rather than the capture device (H_swap_device), and parking of stored Cholesky factors on CPU during recursion so per-expert factors do not accumulate on the device.
Running this quant. EXL3 weights are loaded by exllamav3, for example through tabbyAPI. The vLLM, SGLang and LightLLM commands in the Deployment section below describe the original BF16 release and do not apply to this repository.
Released Resources
Table with columns: Resource, Location, Description| Resource | Location | Description |
|---|
| Model weights | This repository | Weights, config and tokenizer |
| Example trajectories | trajectories/ | Message and tool logs with figures |
| Production skill | GitHub | Workflow, craft guides, rendering tools |
Deployment
The following commands serve the original BF16 release of SenseNova 6.8 Code2Video. They do not load the EXL3 weights in this repository, which run on exllamav3 (see Quantization).
docker run --rm --gpus all --ipc=host -p 8000:8000 \ -v /path/to/SenseNova-6.8-Code2Video:/model \ ghcr.io/modeltc/lightllm:main \ python -m lightllm.server.api_server \ --model_dir /model \ --tp 4 \ --max_req_total_len 524288 \ --host 0.0.0.0 --port 8000
docker run --rm --gpus all --ipc=host -p 8000:8000 \ -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \ -v /path/to/SenseNova-6.8-Code2Video:/model \ vllm/vllm-openai:latest \ --model /model \ --tensor-parallel-size 4 \ --max-model-len 524288 \ --reasoning-parser qwen3 \ --enable-auto-tool-choice \ --tool-call-parser qwen3_coder \ --host 0.0.0.0 --port 8000
docker run --rm --gpus all --ipc=host -p 8000:8000 \ -e SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 \ -v /path/to/SenseNova-6.8-Code2Video:/model \ lmsysorg/sglang:latest \ python -m sglang.launch_server \ --model-path /model \ --tp-size 4 \ --context-length 524288 \ --reasoning-parser qwen3 \ --tool-call-parser qwen3_coder \ --host 0.0.0.0 --port 8000
Recommended harness: Claude Code
A complete film is a long-horizon task: the model writes and renders code, reviews the frames and revises, turn after turn. SenseNova 6.8 Code2Video sustains runs of more than 10 hours and more than 2,000 turns. Configure Claude Code for runs of that length:
Table with columns: Setting, Value, Configuration| Setting | Value | Configuration |
|---|
| Context window | 512K | CLAUDE_CODE_MAX_CONTEXT_TOKENS=524288 |
| Auto-compaction | 0.8 | CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80 |
| Max output | 64K | CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000 |
| Turns per run | 2,000+ | --max-turns 2000 or more |
| Time per run | 10 h+ | timeout 10h or more |
Set the context window explicitly, since Claude Code otherwise treats an unknown model as a 200K model. With compaction at 0.8 and 64K of output, every request stays within the 512K server limit.
Put the endpoint and these variables in a private settings file, and supply the credential through ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY:
{ "env": { "ANTHROPIC_BASE_URL": "https://your-endpoint.example.com", "ANTHROPIC_MODEL": "your-deployed-model-name", "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "524288", "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "80", "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "64000" }, "model": "your-deployed-model-name"}
Avoid slow shell initialization: every Bash call starts a new shell.
Usage
Producing a film has three stages:
Table with columns: Stage, Performed by, Output| Stage | Performed by | Output |
|---|
| 1. Prompt enhancement | A strong general model (we use Claude Opus 5.5) with the Enhance prompt template | dense_prompt.md |
| 2. Production | SenseNova 6.8 Code2Video in Claude Code with the code-video skill | out/final.mp4, SPEC.md, VERIFY.md |
| 3. Revision (optional) | The creator writes notes; SenseNova 6.8 Code2Video applies them in the same project | A new version of the film |
Stage 2 with the settings above:
cd /path/to/task # contains dense_prompt.md, assets/ and .claude/skills/code-videoclaude -p --settings /path/to/private-settings.json --model your-deployed-model-name \ --max-turns 2000 --permission-mode acceptEdits --allowedTools "Bash,Read,Write,Edit,Glob,Grep,Skill" \ 'Use the project code-video skill. Read .claude/skills/code-video/SKILL.md and dense_prompt.md, then the relevant craft guides. Produce the complete video and deliver out/final.mp4, SPEC.md and VERIFY.md.'
The run executes shell commands without asking, so start it in an isolated working environment. For a revision, run claude -p --continue in the same folder with the creator's notes as the prompt.
Copy the skill into .claude/skills/code-video unchanged and put task-specific requirements in the prompt. Setup steps and dependencies are in the GitHub repository.
Production trajectories
trajectories/all.openai.jsonl holds one film per line, in OpenAI messages format.
Each line contains messages, tools, images, enable_thinking and metadata. The messages include reasoning, tool calls and results, retries and compaction summaries. Images appear as <image> placeholders; the ordered images array gives their paths relative to trajectories/, and all 357 figures are in trajectories/figures/. Private paths and hosts are anonymized, so recorded commands are not directly replayable.
License
Released under the Apache License 2.0.
Contributors
We thank the following contributors for their work on SenseNova 6.8 Code2Video: