Qwen3.8-27B: what it is and where it fits

Qwen3.8-27B is a dense vision-language model for conversation, coding, tool use and analysis of text, images and video. Its language component has 27 billion parameters and 64 layers. The official model card declares a native 262,144-token context, a thinking mode and Apache 2.0 licensing.

Those are model capabilities. They do not decide whether the model fits a company's process. We ran a quantised profile on a home RTX 3090 and tested a Go repair, a CSV import and an ATS frontend. The model completed the Go repair, missed one validation rule in the import and produced a frontend that needed repair. That is a more useful description than a feature list alone.

What Qwen3.8-27B means

The 27B label means roughly 27 billion parameters in the language model. It is dense, so the full language component participates in each token computation. Parameter count is not a quality score. It is a meaningful signal for memory, latency and serving cost.

The model card describes 64 language layers. The layout repeats three Gated DeltaNet blocks followed by one Gated Attention block sixteen times. In total, that is 48 DeltaNet blocks and 16 gated full-attention blocks. The model also includes Multi-Token Prediction, or MTP.

Qwen declares a native 262,144-token window and an extension to one million tokens through YaRN. That is a technical limit rather than a latency promise. In our retrieval check, three facts placed near the beginning, middle and end of a synthetic record were returned for inputs of 50,059, 115,074 and 230,085 tokens. The runs took 59.3, 172.6 and 563.4 seconds.

The model has a separate vision encoder, and the card describes image and video input. A company should test its own scans, charts or recordings before making that capability part of a workflow.

What we tested locally

All three tasks used Qwen3.8-27B through Qwen Code. The selected profile used a frozen third-party patched vLLM stack, a 150k limit, W4A16 AutoRound weights, an FP8 KV cache and MTP-3. It was an after-hours home-PC experiment with one active session.

The first frontend rendered 8,652 rows, 251,327 DOM elements and a 485,065-pixel document. After repair, an independent check measured 722 DOM nodes, 15 visible rows, a 900-pixel document and no horizontal overflow. It is a concrete example of the gap between code that looks convincing and code checked in a working browser.

Each task ran once. The figures describe those runs and do not estimate a success rate.

How it fits on a 24 GB card

An RTX 3090 has 24 GB of VRAM. Four-bit weights let our selected profile run on one card. The Go task peaked at 22,539 MiB. Weights are only one part of the memory budget: the context cache, server buffers, CUDA graphs and speculative decoding also need room.

BF16 weights for 27 billion parameters would require about 54 GB before those additional allocations. A 24 GB deployment therefore depends on the weight format, context limit and server settings. A profile handling one request at a time does not provide evidence for team throughput.

The same PC retained an llama.cpp fallback with Dynamic V3 Q4_K_M GGUF weights and a 120k context. The vLLM profile was faster on the matched task, but the weights, cache, engine and context limit all changed. That run does not isolate the quality of one quantisation format.

Tasks to test

Use those conditions to define a Qwen3.8-27B test: one session reads a large repository, uses several tools, works with visual material or maintains a plan across many steps. A local endpoint can keep processing within a chosen environment and makes the marginal token cost predictable after the service is running. Only a measurement on the target process can establish fit.

It may be excessive for short classification, a few extracted fields or brief Polish replies. Compare a smaller model, a deterministic parser and a managed API for those cases. Start with error cost, response-time requirements and data rules. Parameter count narrows the shortlist; it does not make the decision.

What the Codex score means

After the Go task, Codex checked correctness, regression tests, API compatibility, scope, verification and documentation. It awarded 100/100 under Syntalith's rubric. This is a post-run score from the Codex model. An independent laboratory did not assign it. The same local Qwen received 87/100 when driven by the Codex harness and 98/100 in a Claude Code compatibility experiment. The program guiding the model through the repository changed.

Full profiles, timings and failures are in the RTX 3090 experiment results. The source data is available as the run summary, Go-task metrics and long-context check.

If you want to test a local model on your own documents or repository, a free process scan can define the cases, quality threshold and alternatives. The result may be Qwen, a smaller model or an API.