This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

Qwen3.8-27B model inputs

Integrity-checked BF16 model snapshot for self-hosted inference

    Qwen3.8-27B is an upstream Qwen vision-language model with coding, reasoning, and tool-use support. The weights and accompanying files retain their upstream Apache-2.0 license and ownership. This package provides model inputs; a compatible inference runtime and an agent client are supplied separately.

    include.MODULE.bazel owns the immutable upstream revision and SHA-256 integrity values. All 18 BF16 weight shards, the shard index, tokenizer, chat template, model and processor configuration, upstream README, and license come from that same revision. Weight checksums use the publisher’s Hugging Face LFS SHA-256 object identifiers; small files are hashed after verification against their upstream Git blob identifiers.

    • :metadata downloads the configuration, tokenizer, index, and upstream documentation without the weight shards.
    • :weights provides the 18 weight shards.
    • :model provides every file required for the upstream model snapshot.
    • :snapshot creates a flat, uncompressed model archive with original upstream filenames, suitable for extraction into a local model directory.

    Model targets are marked manual to exclude these large downloads from wildcard builds. Build an explicit target to fetch the model:

    bazel build \
  //third_party/co_huggingface_qwen_qwen3_8_27b:snapshot
mkdir -p out/qwen3_8_27b/model
tar -xf bazel-bin/third_party/co_huggingface_qwen_qwen3_8_27b/qwen3.8-27b.tar \
  -C out/qwen3_8_27b/model

    The complete snapshot is about 55.6 GB. Allow additional disk space for the Bazel download cache, archive, and extracted directory. The weight data alone occupies about 55.6 GB (51.8 GiB); inference also needs GPU memory for runtime buffers and the context cache. An 80 GB GPU is a reasonable starting point for one agent, but context length and actual performance need validation on the chosen runtime and GPU. Start with a bounded context such as 32K tokens.

    Point an already provisioned, compatible runtime at the extracted directory and enable its offline mode to prevent fallback downloads. Follow the upstream serving guidance for architecture and tool-call support. This package does not provision a GPU, install a serving runtime, start a service, or change host configuration.