Qwen / Open-weight checkpoint reference

Qwen3.8-27B

Documented in Aguasto Research. This checkpoint has not been evaluated or deployed by Aguasto and is not available through an Aguasto API.

Original Aguasto conceptual illustration. Not manufacturer geometry, an operating Aguasto asset, or a measured benchmark.

Checkpoint identity

Architecture
Dense causal language model with vision encoder; hybrid Gated DeltaNet and Gated AttentionQwen3.8-27B official model card (external source) · Qwen3.8-27B checkpoint configuration (external source) · Qwen3.8-27B Apache 2.0 license (external source)
Total parameters
Not verified for this recordPublished rounded counts; total stored weights determine the preliminary footprint, including all MoE experts.Qwen3.8-27B official model card (external source) · Qwen3.8-27B checkpoint configuration (external source) · Qwen3.8-27B Apache 2.0 license (external source)
Active parameters
Not verified for this recordActive parameters are a compute-related count, not the model's full stored weight count.Qwen3.8-27B official model card (external source) · Qwen3.8-27B checkpoint configuration (external source) · Qwen3.8-27B Apache 2.0 license (external source)
Context
262,144 tokensNative configuration where directly verified. Extensions, advertised K-labels and configuration differences are preserved in the notes below.Qwen3.8-27B official model card (external source) · Qwen3.8-27B checkpoint configuration (external source) · Qwen3.8-27B Apache 2.0 license (external source)
License
Apache 2.0The exact checkpoint includes Apache 2.0 terms. Redistribution, notice preservation, modified-file and patent-license conditions apply. This is distinct from any proprietary Qwen Cloud API offering, which has its own service terms.Qwen3.8-27B official model card (external source) · Qwen3.8-27B checkpoint configuration (external source) · Qwen3.8-27B Apache 2.0 license (external source)

Theoretical weight-memory footprint

A weight estimate is unavailable without a verified parameter count.

This estimate excludes KV cache, activations, quantization metadata, non-weight buffers and serving overhead. It is not a tested GPU fit. MoE routing activates only part of the network per token while the complete expert weights still need storage.

Explore the memory estimator

Quantization

The main checkpoint text configuration uses BF16. The publisher lists a separate FP8 artifact; quantized variants and kernels must be evaluated separately rather than inferred from the family name.

Qwen3.8-27B official model card (external source) · Qwen3.8-27B checkpoint configuration (external source) · Qwen3.8-27B Apache 2.0 license (external source)

Serving considerations

  • The card lists 27 billion language-model parameters and a separate vision encoder. The full combined artifact count is left unknown rather than treating the text-model number as a complete multimodal memory budget.
  • The card and text configuration agree on 262,144 native positions. One-million-token operation uses an extension and is distinct from native context; the proposed hosted API's limits must not replace the open-checkpoint limit.
  • Image/video preprocessing, visual-token expansion and the hybrid attention implementation are part of the deployment boundary. Include visual workload size alongside textual prompt/output lengths when sizing cache and measuring latency.
  • The publisher explicitly documents vLLM and SGLang paths. Thinking is enabled by default and reasoning controls are request-sensitive; freeze request settings and chat templates in any evaluation manifest.
  • Aguasto has not tested or hosted this checkpoint. Documentation does not establish a particular GPU configuration, concurrency, quality level or available service.
Qwen3.8-27B official model card (external source) · Qwen3.8-27B checkpoint configuration (external source) · Qwen3.8-27B Apache 2.0 license (external source)

Framework documentation

These are third-party documentation starting points. Support depends on the exact checkpoint, runtime version and hardware.

Sources & verification

The register preserves each source's publication and verification date. Manufacturer claims and documentation are attributed; Aguasto has not run the cited products in production.

  1. Qwen3.8-27B official model card ↗

    Exact identity and open weights; 27B language-model parameter boundary; Dense vision-language architecture with image/video understanding; 262144 native context versus extended context; Explicit vLLM/SGLang paths and thinking controls. Publication: date not stated in the reviewed source. Verified: 2026-10-13.

  2. Qwen3.8-27B checkpoint configuration ↗

    Qwen3_5ForConditionalGeneration architecture and vision configuration; 262144 text max_position_embeddings; BF16 text dtype; Hybrid attention configuration. Publication: date not stated in the reviewed source. Verified: 2026-10-13.

  3. Qwen3.8-27B Apache 2.0 license ↗

    Apache2.0 checkpoint terms distinct from API service terms. Publication: date not stated in the reviewed source. Verified: 2026-10-13.