isiroai/Qwen2.5-7B-Instruct-TIC

Qwen2.5-7B-Instruct TIC

Bit-exact lossless .tic compression of Qwen/Qwen2.5-7B-Instruct.

Same weights. No quantization.

What is TIC?

Precision	BF16
Raw weights	15.23 GB
.tic weights	10.86 GB (28.7% savings)

Benchmarks

Capacity-primary A/B (eager, RTX 5090): published report benchmarks/qwen2.5-7b-instruct/rtx-5090/report.md.

Methodology: benchmarks README.

Run

# Install
curl -fsSL https://isiro.ai/install.sh | sh

# Download .tic bundle
pip install -U huggingface_hub
hf download isiroai/Qwen2.5-7B-Instruct-TIC --local-dir Qwen2.5-7B-Instruct-TIC

# Serve (vLLM supported today)
isiro serve Qwen2.5-7B-Instruct-TIC --target vllm

Install options · Serve docs · Compile your own

Verify

Bit-exact checks. Hash manifest: model.tic.manifest.json.

# Confirm .tic integrity via the hash manifest
isiro verify Qwen2.5-7B-Instruct-TIC

# Optional: confirm .tic weights match the raw weights (requires compiler)
hf download Qwen/Qwen2.5-7B-Instruct --local-dir Qwen2.5-7B-Instruct
isiro verify Qwen2.5-7B-Instruct-TIC -r Qwen2.5-7B-Instruct

License

Weights: same license as Qwen/Qwen2.5-7B-Instruct (Apache-2.0).

.tic, ISIRO Runtime, and Compiler: ISIRO EULA.