Cadence 0.8B browser research demo
This repository contains the full text-only browser export of Cadence 0.8B, a 758,329,152-parameter typed decision model. It compares supplied answers for choice, yes/no (noul), and ordinal rating questions. It does not generate chat replies.
Use the Cadence demo and its browser implementation. These custom layer graphs and pointer head are not compatible with a generic text-generation pipeline or AutoModel loader.
Download and device limits
The verified artifact contains 90 files / 1,542,947,107 bytes including its manifest. This model card and Hub metadata are additional. The demo downloads about 1.54 GB only after explicit consent, plus approximately 28 MB of browser runtime from the demo site. Inputs are evaluated on-device rather than sent to an inference API. The download host sees ordinary network information including the visitor's IP address.
A desktop browser with substantial available memory is recommended. The tested profile is Chrome on Linux, ONNX Runtime Web 1.30.0, single-thread WASM, FP32 arithmetic. Mobile layouts were inspected; mobile hardware inference and physical WebGPU acceleration are not qualified. Browser execution accepts one typed question, at most 16 answers and 512 packed tokens. Oversized inputs are rejected rather than truncated. The demo verifies every file before caching it; interrupted downloads reuse only completed, verified files. Browser storage can be unavailable or evicted.
Reproducibility and provenance
Manifest SHA-256: 3b773382a7463f5b2bffcba7dd0d0896fbccff33e85ed65cedf60fea19529657.
- Cadence:
foolish-bandit/cadence-0.8b@17cdf36034677e406033fbf97637b4db1984ca70. - Qwen base:
Qwen/Qwen3.5-0.8B-Base@dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68. - Released Kev dependency:
jaredpalmer/kev-0.8b@9a45d25eb2ab761841196625383fa1dff0e56c1e.
This export bundles the reconstructed text backbone, unmerged trained adapter and learned pointer head. Vision is omitted. BF16-representable weights are stored as BF16 bits and losslessly unpacked into FP32; other weights retain their FP32 bits. Computation is FP32 with recurrent DeltaNet lowering. It is not an INT8 model or a CUDA BF16 equivalence claim.
All asset byte counts and SHA-256 hashes appear in manifest.json. The export-time browserParityVerified: false and cudaParityVerified: false fields are conservative generation metadata and have been preserved without modifying the validated artifact. Subsequent, narrowly scoped browser qualification is documented separately below; CUDA parity remains unverified.
Qualification and limitations
Full-stack evidence compares nine real Chrome/WASM fixtures against a separate native CPU FP32 reference, including all three query types and exact 511/512-token cases. All leading answers agree; maximum logit drift is 1.717e-5 and maximum probability drift is 1.969e-6, within the declared 1e-4 / 1e-5 gates. Tokenizer evidence covers 115 exact encoding fixtures. Real local production UI evidence covers actual inference, cache reuse, corruption rejection and recovery from an oversized input.
These checks establish functionality and numerical agreement for the tested cases, not broad task accuracy, cross-device portability, long-context quality or a latency/memory guarantee. Cadence is a research project. Experimental model outputs can be wrong; high output probability is not a guarantee of correctness. Evaluate it for your use case. The original model card describes training/evaluation scope and corpus provenance.
Licensing
Apache-2.0 applies to the model/code contributions with the original upstream attributions preserved. Read LICENSE, NOTICE, LICENSE.QWEN, NOTICE.QWEN and EXPORT_NOTICE. The latter clarifies that original adapter-distribution notices describing separately downloaded upstream weights do not describe this derived export: this browser export bundles the reconstructed text backbone and trained adapter/head. No corpus rows or new dataset license grants are included. No upstream endorsement is implied.