Huihui Flash-Next: experimental Basalt vision overlay

This is a vision-only add-on, not a complete language model. Local assembly and file-integrity checks do not establish GPU runtime compatibility or image understanding quality. GPU validation, 262144-context image tests, and private Hermes image routing remain pending. Do not use private documents to validate it.

The projector is copied byte-for-byte, not requantized. Tokenizer metadata is converted to a metadata-only GGUF. Language tensors are not downloaded, changed, copied or uploaded by this overlay build. Full-pack conversion remains a separate fallback; neither approach automatically replaces the other.

Sources and pins

  • Projector: official Huihui v2 GGUF repository, revision 411f1f7349c9711e6f75aa8d20e1612eb3a4d55d, file mmproj-Qwen3.8-Flash-Next-BF16.gguf, 907543008 bytes, SHA256 b1a82259702816a5330d7bd7607cd9676b11780e79ff7348c21103ff3ce49bd0. This is NOT the Swift projector. Upstream licensing and attribution apply. Its card declares Qwen Community License 1.0, but its LICENSE link was missing when assembled. The included LICENSE is fetched from the official Qwen base; its pinned source revision is recorded in license-source.json.
  • Text model: existing Basalt artifact, revision db74026775af2ee80366d4f6172b16f838b2a933. See manifest.json for its exact size/SHA256 and runtime pins.
  • Basalt source: jesdga95/basalt, commit 463d92e8edf18cf93839bf4cce541ddf7b073280. The loader hook is an experimental integration, not an upstream guarantee.

Contents and restore

manifest.json binds this overlay to the exact text pack's metadata digest, size, image marker IDs and projector geometry. checksums.json records all published file sizes/SHA256 values. The projector and tokenizer GGUF must remain beside the manifest. No credentials, user prompts or private session data belong in this repository.

  1. Restore and checksum-verify the existing text model and saved CUDA13/sm120 runtime on the GPU host, using the base artifact's restore instructions.
  2. Download this repository at a pinned commit into a separate directory. Verify checksums.json before installation. Keep the base model untouched.
  3. Run python install-vision-overlay.py --source /workspace/basalt-test/source. This backs up the Python loader and installs an opt-in hook only.
  4. Start a separate test configuration with server.vision="dynamic", 262144 context, INT8 KV and spec3. Set BASALT_VISION_OVERLAY_MANIFEST=/absolute/path/to/manifest.json for that process only. The dynamic mode does not require the optional vision sidecar.
  5. Verify synthetic image content via the authenticated private API; then check text/tools, long-context resource use and Hermes native-image routing.

Rollback: unset the environment variable and use the original text configuration. The installer does not start a server, alter the pack, provision a GPU, or publish an endpoint. Missing files, incorrect metadata, hashes or marker IDs fail closed. The base pack must still be independently SHA256-verified during restore; metadata matching alone does not validate every language tensor.

Downloads last month
67
GGUF
Model size
0 params
Architecture
qwen4exp
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support