The useful question after an open-weight release is what you can run, at which precision and context length. “27B” alone cannot answer whether your GPU has enough usable memory.
What is released
The official Qwen3.8-27B model card lists Apache 2.0 weights and a multimodal architecture with hybrid attention. Its native context is 262,144 tokens. Open weights allow inspection and deployment under the license; they do not mean that the complete training data and training procedure were released.
The card enables thinking by default. Output behavior therefore depends on the chat template and serving parameters, not just the checkpoint name. Vendor benchmark scores describe the stated evaluation setup; they are not our independent results or a forecast for every coding project.
A weight-only calculation
For a simplified set of exactly 27 billion parameters, multiply parameter count by stored bytes per parameter. At 16 bits, that is 54 GB, approximately 50.3 GiB. At eight bits, it is 27 GB or 25.1 GiB; at four bits, 13.5 GB or 12.6 GiB.
These cover-image values are arithmetic illustrations. They exclude extra tensors, quantization scales, vision components, activations, attention state, temporary buffers and framework allocations. GB uses powers of 1,000; GiB uses powers of 1,024. Comparing “13.5 GB” directly with a card's advertised capacity and declaring that serving fits skips most of the problem.
The serving recipe supplies the missing context
The vLLM recipe documents materially different configurations. Its single RTX 5090 NVFP4 example uses a 32,768-token maximum and eager execution; the two-card examples use different settings. Those are recipe authors' validated configurations, not measurements by PeopleAreGeek. A third-party quantization is also a different artifact from the official FP8 checkpoint.
For a reproducible trial, record checkpoint revision, quantization source, engine version, usable device memory, maximum context and simultaneous sequences. Start with a representative short request, then a long document and concurrent requests. Check correctness and completed task time alongside peak memory.
A server that starts with one short prompt may still fail under a larger batch. Conversely, reducing the maximum context can make a configuration practical without changing the underlying model. Keep these deployment tradeoffs explicit rather than promising local operation on any machine or treating a benchmark rank as a hardware recommendation.
Separate open weights from complete open training; calculate weight storage rather than promise device fit; compare actual serving recipes and context overhead.