Downloads · 30 days
15
33% of all-time downloads
trevor000/Qwen3.8-27B-Escha-W2-RTX4070
Qwen3.8-27B-Escha-W2-RTX4070 is a machine learning model from trevor000. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
This is an attributed mirror of EschaLabs/Qwen3.8-27B-Escha-W2, prepared for reproducible RTX 4070 research. The weights are unchanged upstream files; this project did not train them. opencode.json is intentionally ex…
Downloads · 30 days
15
33% of all-time downloads
All-time downloads
45
Public
Parameters
6.3B
10.2 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors10.2 GB · 100%
How the weights are stored.
I163.8B · 59%
From the Hugging Face model README
This is an attributed mirror of EschaLabs/Qwen3.8-27B-Escha-W2, prepared for reproducible RTX 4070 research. The weights are unchanged upstream files; this project did not train them. opencode.json is intentionally excluded from the publication allowlist.
The manifest allowlists the two safetensors weight shards, the safetensors index, configuration, generation settings, tokenizer files, vocabulary/merges, chat template, and quantization metadata from /home/trevor/inference-lab/models/escha-w2. The model’s LICENSE and THIRD_PARTY_LICENSES/Qwen-LICENSE.txt are included here as notices. No runtime binary or custom kernel is bundled; the separate Escha runtime model repository is named in provenance.json.
Download the reviewed files with hf download trevor000/Qwen3.8-27B-Escha-W2-RTX4070 --local-dir Qwen3.8-27B-Escha-W2-RTX4070. The model source revision is 561dae0aeea1efbdafe122f2e385fdeaa99127e7; use the runtime's own documented launcher and license rather than assuming this mirror is a standalone Transformers package.
Historical lab context includes a 148.03 tok/s median with a deliberately tiny 176-token cache control, plus a usable product fallback around 125 tok/s at 5,120 exact active tokens with a 12,288-token rolling RAM archive. Those measurements require the lab's modified SGLang runtime and separate Apathy DFlash draft, not these weights alone. See the measured hill-climb report and chat launcher. Archive recall is not equivalent to full exact attention over the archive. This mirror adds no new GPU benchmark.
The card declares Apache-2.0 and identifies Qwen/Qwen3.8-27B as the base. Retain the included Qwen third-party license and Escha Labs license. This mirror has no upstream endorsement; cite Escha Labs and Qwen, and review the separate Escha runtime license before redistributing runtime code.