Short answer: Qwen3.8-27B is a 27B dense model that fits entirely on a single consumer GPU, no offloading tricks required. The 4-bit GGUF quant needs about 14.3-17.6 GB, comfortable on a 24GB RTX 4090 or 5090 with headroom for context. That makes it the easiest genuinely capable local model to run this month, especially next to 300B+ parameter MoE releases like GLM-5.2 and DeepSeek V4 Flash that demand 80GB or more just to start. This guide covers the exact memory numbers and which GPU tier gets you what.
Quick Navigation:
Qwen3.8-27B Architecture
Qwen3.8-27B landed in mid-August 2026 and spread fast: a 375-point, 177-comment Hacker News thread formed around its Artificial Analysis benchmark score within days, alongside a separate 299-point thread on the release announcement. It ships Apache 2.0 licensed with open weights, and both text and vision-language (image and video) input.
Qwen3.8-27B
Unlike the MoE models that have dominated recent open-weight releases, Qwen3.8-27B is dense: every one of its 27B parameters activates on every token, using a hybrid pattern of Gated DeltaNet (linear attention) layers mixed with full Gated Attention layers across 64 total layers. That’s the architectural reason it handles long context efficiently without the memory overhead a pure quadratic-attention model would carry at 262K+ tokens. For the difference between dense and MoE memory math, see our LLM quantization guide.
Memory Requirements by Quantization
These figures are Unsloth’s dynamic GGUF quantizations, live on Hugging Face the same day as release.
| Quantization | File Size | Realistic Hardware |
|---|---|---|
| 1-bit (UD-IQ1_S/M) | 6.2-6.7 GB | 8GB GPU (laptops, entry cards) |
| 2-bit (UD-Q2_K_XL) | 7.3-9.8 GB | 12GB GPU (RTX 3060, 4070) |
| 3-bit (UD-Q3_K_XL) | 10.9-13.1 GB | 16GB GPU (RTX 4060 Ti, 4080) |
| 4-bit (UD-Q4_K_XL) | 14.3-17.6 GB | 24GB GPU (RTX 4090, 5090) |
| 6-bit (UD-Q6_K_XL) | 22-25.3 GB | 24-32GB GPU (4090, 5090) |
| 8-bit (Q8_0/UD-Q8_K_XL) | 29-31.5 GB | 32GB+ GPU (RTX 5090, A6000) |
| 16-bit (BF16) | 54.7 GB | 2x 24GB GPUs or 1x 80GB card |
Key insight: every quant level from 1-bit through 8-bit fits on a single GPU people already own. Compare that to DeepSeek V4 Flash’s 82.5 GB floor or GLM-5.2’s 223 GB floor, both MoE models that need CPU offloading or multiple GPUs just to load. Qwen3.8-27B is the first release in months where “which GPU do I already have” is a real answer, not just an entry point to a bigger build.
Four Ways to Run Qwen3.8-27B
Path 1: Budget 12-16GB GPU (2-bit to 3-bit)
A 12GB card like the RTX 3060 or 4070 handles the 2-bit quant at 7.3-9.8 GB, and a 16GB RTX 4060 Ti or 4080 comfortably runs 3-bit at 10.9-13.1 GB. This is a genuinely low barrier to entry, no offloading, no second GPU, just a mid-range card most enthusiasts already own.
Path 2: 24GB GPU, the Sweet Spot (4-bit)
An RTX 4090 or 5090 running the 4-bit quant at 14.3-17.6 GB leaves several GB free for a long context window, close to the model’s full 262K native limit. Community-reported throughput on a 24GB card was around 50 tokens/sec at 256K context within days of release, though that’s an independent benchmark, not an official number. See our RTX 5090 vs 4090 comparison for which card fits your budget.
Path 3: 32GB+ GPU, Near-Lossless (8-bit)
A 32GB RTX 5090 or a workstation card like the A6000 fits the 29-31.5 GB 8-bit quant with almost no quality loss versus the full BF16 weights, and still leaves room for a solid context window. This is the setup if you want maximum quality without going to a second GPU.
Path 4: Rent Cloud GPUs (Best for Trying It First)
A single rented 24GB card is enough to test 4-bit before buying anything, and it’s a cheap test given how little VRAM the whole useful range needs. Vast.ai usually has the lowest hourly prices, and RunPod gives new users a $5 credit to test with. See our cloud GPU comparison for current pricing.
Qwen3.8-27B vs GLM-5.2 vs DeepSeek V4 Flash: Which Actually Fits Your GPU?
All three made headlines this summer, but they sit on opposite ends of the self-hosting difficulty curve. GLM-5.2 and DeepSeek V4 Flash are 300B+ parameter MoE models that need 82.5-223 GB of combined memory even at their smallest usable quant, forcing CPU offloading, a Mac Studio with 128GB+ unified memory, or multiple GPUs. Qwen3.8-27B is dense and an order of magnitude smaller: its entire practical range, from 1-bit to 8-bit, fits inside 32 GB, on a single GPU, with no offloading at all.
That tradeoff cuts both ways. Qwen3.8-27B isn’t competing to beat 300B-class models on raw benchmark ceiling, it’s competing on being genuinely easy to run well. If your priority is “what can I load on the GPU I already own today,” Qwen3.8-27B wins by a wide margin. If you want the highest possible capability ceiling and have the hardware for it, our DeepSeek V4 Flash guide and GLM-5.2 guide cover what those bigger builds require.
Pro tip: if you’re running the 4-bit quant and want more context headroom, quantize the KV cache with –cache-type-k q4_1 –cache-type-v q4_1 in llama.cpp. It roughly triples usable context at the same VRAM budget, useful given Qwen3.8-27B’s 262K native window.
FAQ
Can you run Qwen3.8-27B on a single GPU?
Yes, comfortably. Unlike the 300B+ MoE models dominating recent releases, Qwen3.8-27B is a 27B dense model. The 4-bit GGUF quant needs roughly 14.3-17.6 GB, which fits a single RTX 4090 or 5090 with room to spare for context. No CPU offloading or multi-GPU setup required.
How much VRAM does Qwen3.8-27B need?
It ranges from about 6.2 GB at 1-bit up to 54.7 GB for the full BF16 weights. The practical sweet spot is 4-bit at 14.3-17.6 GB (fits a 24GB GPU) or 8-bit at 29-31.5 GB (fits a 32GB+ card like the RTX 5090) for near-lossless quality.
What’s the cheapest way to run Qwen3.8-27B locally?
A single 12GB or 16GB GPU running the 2-bit or 3-bit Unsloth GGUF quant, which needs 7.3-13.1 GB. That’s a genuinely budget-friendly entry point, something recent 300B-class MoE releases like GLM-5.2 and DeepSeek V4 Flash can’t offer.
Is Qwen3.8-27B better than DeepSeek V4 Flash or GLM-5.2 for local use?
For most home setups, yes, specifically because it’s easier to run. DeepSeek V4 Flash and GLM-5.2 are 300B+ parameter MoE models that need 80GB+ of combined memory even at their smallest usable quant. Qwen3.8-27B’s entire useful range fits on GPUs people already own. Whether it’s better on raw capability depends on the task; it’s a smaller model competing on efficiency, not on beating 300B-class benchmarks outright.
Why is Qwen3.8-27B getting so much attention?
It’s being called the best local dense model in its size class, backed by a 375-point Hacker News thread over its Artificial Analysis benchmark score. Community reports followed fast: 50 tokens/sec at 256K context on a 24GB GPU, 200 tok/s on an RTX 5090, and working setups on a plain RTX 3090, all within days of release. Apache 2.0 licensing and Unsloth GGUF quants landing same-day helped it spread quickly.
Related Reading
DeepSeek V4 Flash Hardware Guide
The 304B MoE model that needs 82.5GB+ to start
GLM-5.2 Hardware Guide
The 744B model that needs 245GB+ just to start
LLM Quantization Guide
GGUF, GPTQ, AWQ and what the bit levels really cost
RTX 5090 vs 4090 for Deep Learning
Which 24GB+ card fits a model this size
Multi-GPU Ollama Setup
For when even a single 32GB card isn’t enough
Building for Local LLMs?
Want a build that can grow into bigger models later? Check our Tailored Builds page.