Qwen3.8 Guide: VRAM Requirements, GGUF Builds

September 26, 2026September 28, 2026

About This Model

Qwen3.8 is the series the Qwen team calls “the most capable generation in the Qwen open-model family to date." Its core model, Qwen3.8-27B, is a 27B dense model built as a native vision-language model (VLM) that understands images and video. Built on the Qwen3.5 architecture, it is said to bring substantial gains in coding, professional work, research and long-horizon agentic tasks.

It uses a hybrid layout: of its 64 layers, 48 use linear attention (Gated DeltaNet) and 16 use standard attention (Gated Attention). The native context is 262,144 tokens, extensible to one million tokens with YaRN. Thinking is on by default; it can be disabled per request, and reasoning_effort sets its depth to xhigh, medium or low.

What Makes It Stand Out

1. Agentic coding improved sharply over the previous model of the same size. An excerpt from the publisher’s comparison tables (bold marks the best value in each row):

• It beats the previous Qwen3.6-27B on every row in the model card. DeepSWE 1.1 more than tripled, from 13.3 to 42.2.
• It is strongest at tasks where the model has to act and carry the work through to the end. It has the highest score among the compared models on SWE-bench Pro (resolving real GitHub issues), OSWorld-Verified (operating a real OS through its screen) and WebArena-Verified (browser use). On OSWorld-Verified it leads Qwen3.7-Plus and Opus4.6 Max by more than 11 points.

2. It also improved on tasks that involve working from images and video. Vision2Web (building web pages from visuals) rose to 62.9 (previous: 45.0) and SWE-MM (software engineering with images) to 38.6 (previous: 25.7), both the best among the compared models.

Caveats and Weaknesses

• It does not reach the top models on hard knowledge and reasoning questions. It scores below Qwen3.7-Plus and Opus4.6 Max on GPQA Diamond, and trails Opus4.6 Max by 9.2 points on HLE. On Terminal Bench 2.1 it is also 5.2 points behind Opus4.6 Max.
• On visual reading and perception, it is level with or slightly behind Qwen3.7-Plus. Document reading (OmniDocBench 1.5: 91.1 vs 91.4) and real-world perception (RealWorldQA: 85.9 vs 86.9) are within one point, but embodied reasoning (ERQA: 65.5 vs 69.8) is 4.3 points lower.
• Not all scores were measured under the same conditions. For SWE-bench Pro, only Opus4.6 Max uses its officially reported score; the other models were re-evaluated with the Claude Code harness. QwenSWEBench, CoWorkBench and RecreationBench are Qwen’s in-house benchmarks and cannot be reproduced by third parties under the same conditions.

Running It Locally

• The publisher releases the weights in Transformers format. The model card points to Transformers, vLLM, SGLang and TokenSpeed and does not mention llama.cpp, Ollama or LM Studio. GGUF and MLX builds for those tools come from third parties such as unsloth and lmstudio-community. See the memory requirements table below for which quantization fits which GPU.
• Thinking is on by default, so outputs are long. For agentic work, the model card recommends allowing up to 262,144 tokens for reasoning and up to 131,072 tokens for the final response. It also warns that in multi-turn tasks, a lower reasoning_effort does not always finish sooner, because shallower analysis can lead to more failures and retries.
• Enable the one-million-token extension (YaRN) only when you need it. Major engines implement static YaRN, whose scaling factor stays fixed, so leaving it on may hurt performance on shorter texts.
• The license is Apache-2.0, which allows commercial use.
• Articles in this family also cover third-party fine-tunes built on Qwen3.8-27B (Qwopus3.8, Signal-3.8, TWIN-TURBO, Salience and others) and Ternary-Bonsai-2, a ternary rebuild. Their weights differ from the official model, so they are separate models; the scores above are for the official Qwen3.8-27B.

Source: Qwen/Qwen3.8-27B model card, as of 2026-09-26. Benchmark figures are as published by the Qwen team.

Our Coverage and Data

Everything Local Model Watch has published about the Qwen3.8 family: 6 article(s) covering the base model and its fine-tunes, plus converted builds we tracked after publication. Memory requirements below are computed by this site from file sizes, not quoted from model cards. Part of our model family index.

At a Glance

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 27.8B parameters

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build unsloth/Qwen3.8-27B-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but unsloth/Qwen3.8-27B-GGUF provides a GGUF build you can use.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compression: the IQ4_XS build measures 4.11 bits per weight — about 26% the size of the original 16-bit weights, calculated by this site from the actual file sizes.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Quantized and Converted Variants

File sizes of each build:

• Available builds in unsloth/Qwen3.8-27B-GGUF: IQ1_S 5.8GB / IQ1_M 6.3GB / IQ2_XXS 6.8GB / IQ2_S 7.8GB / Q2_K_XL 9.2GB / IQ3_XXS 10.2GB / IQ3_S 11.2GB / Q3_K_XL 12.2GB / IQ4_XS 13.3GB / Q4_K_S 14.3GB / Q4_0 15.0GB / Q4_K_M 15.3GB / Q4_1 16.3GB / Q4_K_XL 16.4GB / Q5_K_S 17.4GB / Q5_K_M 18.4GB / Q5_K_XL 19.4GB / Q6_K 20.5GB / Q6_K_M 21.5GB / Q6_K_L 22.5GB / Q6_K_XL 23.6GB / Q8_K_L 26.1GB / Q8_0 27.1GB / Q8_K_XL 29.3GB / BF16 50.9GB
• Available builds in lmstudio-community/Qwen3.8-27B-MLX-4bit: MLX 4bit 15.0GB
• Available builds in lmstudio-community/Qwen3.8-27B-MLX-8bit: MLX 8bit 27.5GB
• Available builds in unsloth/Qwen3.8-27B-NVFP4: NVFP4 21.0GB
• Available builds in lmstudio-community/Qwen3.8-27B-GGUF: Q4_K_M 15.7GB / Q6_K 20.9GB / Q8_0 27.1GB
• Available builds in prism-ml/Ternary-Bonsai-2-27B-mlx-2bit: MLX 2bit 8.0GB
• Available builds in prism-ml/Ternary-Bonsai-2-27B-gguf-dev: Q2_0 7.1GB

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files.

Articles (the family’s own models first, then newest)

Repositories

• Qwen/Qwen3.8-27B
• prism-ml/Ternary-Bonsai-2-27B-gguf
• bartowski/vectionlabs_Salience-27B-R6-GGUF
• DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NM-DAU-NEO-MTP-GGUF
• agentionai/Signal-3.8-27B-GGUF
• Jackrong/Qwopus3.8-27B-Flash-GGUF

Last updated 2026-09-28 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log.