Qwen3.8-27B Multimodal Vision-Language Model: 8GB+ VRAM, GGUF Builds

September 26, 2026September 27, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

The Qwen team has released Qwen3.8-27B, a 27B dense model on Hugging Face, serving as the core of the new generation of Qwen open models, Qwen3.8. The model card positions Qwen3.8 as “the highest-performing generation of Qwen open models to date," building on the Qwen3.5 architecture with significant improvements in coding, business tasks, research, and long-horizon multi-step agent workflows. It is a vision-language model (VLM) capable of understanding images and videos, allowing users to toggle the presence and depth of thinking per request.

Quantized versions and fine-tunes based on this model are being published in rapid succession. While our site has previously covered 5 derivative versions, this is our first article on the original base model.

Specifications

• Parameter count: 27B (language model portion; has a separate visual encoder)
• Architecture: Dense. A hybrid configuration of 64 layers stacking 16 repetitions of a 4-layer block: three “Gated DeltaNet → FFN" layers followed by one “Gated Attention → FFN" layer. In total, 48 of the 64 layers are linear attention (Gated DeltaNet) and 16 are standard attention (Gated Attention)
• Gated DeltaNet: Linear attention heads have 48 V and 16 QK, with a head dimension of 128
• Gated Attention: Attention heads have 24 Q and 4 KV, with a head dimension of 256 and a RoPE dimension of 64
• Hidden layer dimension: 5,120. FFN intermediate dimension: 17,408
• Vocabulary (embedding/output): 248,320 (including padding)
• MTP (Multi-Token Prediction): Trained for multi-step prediction
• Context length: Native 262,144 tokens. Expandable up to 1,000,000 tokens using YaRN
• Inputs: Text, images, and videos (Causal Language Model with a visual encoder)

Performance

The model card provides two comparison tables for text and vision (VL) respectively. The comparison targets are four models from the previous or competing generations: Qwen3.6-27B, Qwen3.7-Plus, Muse Glimmer-30B, and Opus4.6 Max, with all figures published by their respective creators (“–" indicates no result or not applicable).

Text Performance (transcribed directly from the model card table)

According to the model card, it outperforms the previous-generation Qwen3.6-27B across all rows. The largest gains are seen in agentic coding, where DeepSWE 1.1 surged more than threefold from 13.3 to 42.2, and the creator’s internal benchmark QwenSWEBench rose from 49.3 to 79.0. SWE-bench Pro (the percentage of real GitHub issues resolved by editing repositories) stands at 61.7, the highest among the comparison targets. On the other hand, Terminal Bench 2.1, which measures terminal task completion, scores 73.0, falling 5.2 points short of Opus4.6 Max (78.2). PhD-level scientific reasoning (GPQA Diamond: 89.2) and expert-crafted ultra-hard benchmark HLE (30.8) trail both Qwen3.7-Plus and Opus4.6 Max, with a 9.2-point gap against Opus4.6 Max in HLE. The trend gathered from the table is that it is strong in hands-on, execution-driven coding and business tasks, while falling short of top-tier models in difficult knowledge and reasoning benchmarks. LiveCodeBench v6, which compiles recent competitive programming problems, achieved a top score of 90.3 among the evaluated models.

However, comparison conditions require caution. According to the model card notes, SWE-bench Pro values were officially reported only for Opus4.6 Max, while other models were re-evaluated using the Claude Code harness. QwenSWEBench and CoWorkBench are internal benchmarks by the Qwen team and cannot be independently reproduced by third parties under identical conditions.

Visual (VL) Performance (transcribed from the model card table)

The visual domain shows a similar trend, with significant gains in screen-interaction tasks. OSWorld-Verified, which measures task completion via real OS screen operation, scored 84.3, outperforming Qwen3.7-Plus (73.3) and Opus4.6 Max (72.7) by over 11 points. Browser operation WebArena-Verified (64.8), image-based software engineering SWE-MM (38.6), and web page generation from visual input Vision2Web (62.9) also achieved the highest scores among comparison targets. On the other hand, it trails Qwen3.7-Plus in document reading (OmniDocBench 1.5: 91.1 vs 91.4), real-world photo recognition (RealWorldQA: 85.9 vs 86.9), embodied reasoning (ERQA: 65.5 vs 69.8), and multimodal tool use average score (ClawEval-MM: 56.9 vs 60.1). While the first two are narrow margins within 1 point, ERQA shows a 4.3-point gap. Note that values labeled “With CI" for MathVision, BabyVision, and CharXiv include tools like code execution and differ in conditions from “Without CI"; comparisons must be made between values under identical conditions.

Strengths and Use Cases

The enhancements highlighted in the model card focus on four areas: coding, business tasks, research, and long-horizon multi-step agent workflows, emphasizing autonomous planning and improved capability to sustain tasks based on environmental feedback (execution results and errors). As shown in the tables, notable improvements are concentrated in “execution-to-completion" tasks such as repository editing to resolve issues, as well as operating terminals, browsers, and OS screens, making it best suited as the brain for coding and PC-operating agents. The model card also highlights expanded compatibility with popular harnesses and development tools.

On the visual side, inputs ranging from STEM diagrams and documents to hour-long videos are supported. For handling longer videos, the model card recommends increasing longest_edge in the distributed video_preprocessor_config.json to 469,762,048 (equivalent to 224k video tokens) for finer-grained frame sampling.

Thinking control is managed through the following three mechanisms:

• Thinking mode is enabled by default. It can be turned off per request via enable_thinking: False in chat_template_kwargs.
• reasoning_effort selects thinking depth from xhigh (default), medium, and low. However, the model card warns that in multi-turn agent workflows, lowering the depth does not necessarily reduce total execution time (shallower analysis can increase failures and retries, consuming more tokens instead).
• preserve_thinking is enabled by default, retaining thinking blocks across the entire conversation history. This maintains agent decision consistency and facilitates KV cache reuse. Set to False if you only want to preserve thinking from the most recent utterance.

Recommended sampling settings are temperature=1.0, top_p=0.95, top_k=20, and presence_penalty=0.0 in thinking mode, and temperature=0.7, top_p=0.80, top_k=20, and presence_penalty=1.5 in non-thinking mode. If repetitive text loops occur, presence_penalty can be increased within the 0–2 range, though raising it too high may cause language mixing or a slight performance drop. For agent workflows, allocating sufficient output length is recommended, with target guidelines of up to 262,144 tokens for the thinking portion and up to 131,072 tokens for the final response.

How It Differs from Similar Models

Our site has previously covered derivatives based on this model. For instance, Ternary-Bonsai-2-27B-gguf Text Generation Model: 8GB+ VRAM reconstructs Qwen3.8-27B into a ternary representation at an effective 1.72 bits per weight, aiming to run on modest hardware while minimizing performance loss compared to FP16. Qwen3.8-27B TWIN-TURBO Uncensored GGUF Released is a personal fine-tune adjusted to reduce thinking tokens. Both derivatives trade off size or thinking length based on the base model’s capabilities. The original Qwen3.8-27B serves as the reference point delivering the performance shown in the tables under the creator’s conditions.

Hardware Requirements

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

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-09-27): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.

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.

How to Get It

• Distribution format: Hugging Face’s Qwen/Qwen3.8-27B hosts the post-trained model weights and configuration files in Transformers format (safetensors). Agreement to terms of use is not required for download.
• Example download command: huggingface-cli download Qwen/Qwen3.8-27B
• Supported engines: The model card lists usage with Transformers, vLLM, SGLang, and TokenSpeed. For production deployment or scenarios requiring high throughput, dedicated servers like SGLang, vLLM, and TokenSpeed are recommended. Note that inference efficiency varies significantly across frameworks, so using the latest versions is advised.
• API usage: Text, images (image_url), and videos (video_url) can be passed via OpenAI-compatible Chat Completions APIs to servers set up with the above engines. Currently, vLLM is reportedly the only framework allowing per-request adjustment of video frame extraction (fps, etc.).
• Long-context handling: YaRN must be used for contexts exceeding 262,144 tokens. vLLM, SGLang, and TokenSpeed support rewriting rope_parameters in text_config within config.json or overriding via startup arguments. However, because all frameworks implement static YaRN (constant scaling factor regardless of input length), performance may degrade on short texts. The model card recommends enabling it only when long context is required and adjusting the scaling factor (factor) to match the intended length (e.g., 2.0 for 524,288 tokens).
• llama.cpp, Ollama, and LM Studio are not mentioned in the model card. GGUF versions for these platforms are primarily converted and adjusted by third parties as seen in derivative articles, so verifying the source and quantization method prior to use is recommended.

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

In addition, 29 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

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. What the quantization names mean: glossary.

Related Articles

• Ternary-Bonsai-2-27B-gguf Text Generation Model: 8GB+ VRAM
• Qwen3.8-27B TWIN-TURBO Uncensored GGUF Released
• Signal-3.8-27B-GGUF Token-Efficient Optimized GGUF Model: 16GB+ VRAM
• vectionlabs_Salience-27B-R6-GGUF Vision-Language Model: 12GB+ VRAM

What to Read Next

• Find models by VRAM (This model runs from the 8GB tier) → Other models that run on a 8GB GPU
• Explore the same model family → Qwen3.8 family overview (6 articles, 8 converted builds)
• Engines that run this model → llama.cpp / Ollama / vLLM
• What IQ1_M, IQ4_XS, Q8_K_L mean and where to get this model → GGUF format guide and models / MLX format guide and models / FP8 format guide and models / NVFP4 / MXFP4 format guide and models / Quantization and model-format glossary
• Learn about the publisher → Alibaba (Qwen): models, licenses and articles

Sources

• https://huggingface.co/Qwen/Qwen3.8-27B

Update History

• 2026-09-26: Added converted builds to “Quantized and Converted Variants”: unsloth/Qwen3.8-27B-GGUF, Qwen/Qwen3.8-27B-FP8, lmstudio-community/Qwen3.8-27B-MLX-4bit, lmstudio-community/Qwen3.8-27B-MLX-8bit, unsloth/Qwen3.8-27B-NVFP4, lmstudio-community/Qwen3.8-27B-GGUF
• 2026-09-26: Updated the hardware requirements table with the actual file sizes of unsloth/Qwen3.8-27B-GGUF.
• 2026-09-26: Changed the title to show what the article covers (VRAM requirements, file list, etc.).