Qwen3.8-27B-BF16-eole

Pre-converted Eole checkpoint of Qwen/Qwen3.8-27B. Use Eole to load this repository; its config and tensor names are not a Transformers checkpoint. Weights retain their existing local converted values; this packaging step does not requantize or normalize them. Floating-point weights are stored in BF16.

Use

Install Eole 0.6.4 with the optional dependencies required by this architecture. The tokenizer and runtime configuration are bundled. Paths use ${MODEL_PATH} so the checkpoint can be moved. See the Eole recipes for model-specific prompts and serving configurations.

model_path: eole-nlp/Qwen3.8-27B-BF16-eole
compute_dtype: bf16

For reproducible deployments, download a pinned Hub commit with huggingface_hub.snapshot_download and pass its returned directory as model_path.

Provenance

  • Upstream repository: Qwen/Qwen3.8-27B.
  • Pinned upstream revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
  • Converted from the pinned upstream revision for publication, using the Eole source revision listed below and eight output shards. Converter tensor completeness and transformed-value equality checks passed.
  • Eole source revision: 0c9b3f7ea253b1b977ef47a01ef2a8c6b130f22b.
  • Runtime precision: BF16 for floating-point inference; Eole's FP32 normalization behavior is unchanged.
  • See model-manifest.json for tensor statistics and file SHA256 digests, and UPSTREAM_README.md plus the included license for upstream documentation and attribution.

Validation

Both the portable local package and a fresh private-Hub download passed offline inference smoke tests. All downloaded file SHA256 digests matched the published manifest. See validation.json for device, precision and smoke outputs. These checks are functional validation, not benchmark/quality evaluation. This artifact is intended for inference, not as a claim of fresh benchmark evaluation.

Downloads last month
41
Safetensors
Model size
27B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for eole-nlp/Qwen3.8-27B-BF16-eole

Base model

Qwen/Qwen3.8-27B
Finetuned
(969)
this model