Edge0-35B-A3B-preview — rust-model-inference GGUF
This repository contains a losslessly converted GGUF for the Edge0-35B-A3B-preview text model.
The file is intended for the edge0 implementation in
rust-model-inference,
which keeps Edge0 as its own model architecture and its own MoE path. It is
not a general-purpose llama.cpp GGUF.
File
| File | Size | SHA-256 |
|---|---|---|
Edge0-35B-A3B-preview-lossless.gguf | 19,562,215,008 bytes | 50c6c1ce5faef36d5e72d565fa4a27a04801aa243a0d4a5c3f5c4337a408ec7d |
The conversion preserves the original tensor payloads; it does not apply an
additional quantization step. The GGUF metadata declares
general.architecture=edge0. The base tensors and bundled LoRA tensors are
included in the file.
Scope and limitations
- Text generation only. The source directory contains no visual weights.
- The Rust path uses the model's actual router. The official Edge0 prerouter is not included in this GGUF and is not required to run it.
- Use a revision of
rust-model-inferencecontaining Edge0 support (PR #122 or a later merged revision). - This file is not interchangeable with Qwen3.5 GGUF files and should not be loaded through a Qwen3.5 model entry point.
Download
hf download EvoAwaken-Workshop/Edge0-35B-A3B-preview-gguf \
Edge0-35B-A3B-preview-lossless.gguf \
--repo-type model \
--local-dir ./Edge0-35B-A3B-preview-gguf
Verify the download:
shasum -a 256 \
./Edge0-35B-A3B-preview-gguf/Edge0-35B-A3B-preview-lossless.gguf
The digest must be:
50c6c1ce5faef36d5e72d565fa4a27a04801aa243a0d4a5c3f5c4337a408ec7d
Run with rust-model-inference
The following fetches the Edge0 implementation from PR #122. If the change is already merged, use the repository's default branch instead.
git clone https://github.com/Liyulingyue/rust-model-inference.git
cd rust-model-inference
git fetch origin pull/122/head:edge0-support
git checkout edge0-support
cargo run --release --bin rust-model-inference -- \
--model ./Edge0-35B-A3B-preview-gguf/Edge0-35B-A3B-preview-lossless.gguf \
--prompt "Hello" \
--max-tokens 32 \
--temp 0 \
--threads 1 \
--kv-cache f32 \
--prefill-batch-size 1
Verification
The scalar CPU path was compared with the Edge0 reference implementation with
acceleration paths disabled. For a fixed text sample, 5,316,608 F32 values
matched bit-for-bit across 16 forward passes, including all 40 layer outputs
and final logits. The four greedy token IDs for Hello were:
[9419, 0, 2500, 628]
The verification does not use BLAS, MLX, SIMD, or FMA acceleration for the compared values. Normal inference may use the runtime's available acceleration paths; those paths are not the bitwise verification mode.
Upstream
- Original model: Edge0-35B-A3B-preview
- Edge0 reference commit:
fb4cd2c49ebe22bb230e1451ecb8fb4957ca62e6 - Rust implementation: rust-model-inference
- Upstream license: Apache-2.0