Can I run Qwen3.8-27B on an RX 9060 XT 16GB?
Key facts
No: it needs about 18 GB and this card has 16 GB of graphics memory.
It needs about 17.8 GB at standard quality (estimate). Checked . unsloth UD-Q4_K_M GGUF.
Updated · Qwen3.8-27B on every card
How much memory it needs
That is about 1.2 times what this card has free.
Get a 1-line alert when a faster quant for Qwen3.8-27B drops — or when GPUs that run it drop >10%
Get a 1-line alert when a new open-weight model fits 24 GB VRAM (17.8 GB at Q4 for Qwen3.8-27B) or when used RTX 3090 / 4090 & Mac Studio prices drop >10% — weekly, zero spam.
What runs Qwen3.8-27B instead
The cheapest graphics card on our list that holds the whole model (it still needs the rest of a PC):
GeForce RTX 3090 24GB$1,395€1,246used (median of sold listings)
Fast~43 tok/sAnswers appear faster than you can read them.EU shops · Shown to visitors in Europe.US shops · Shown to visitors outside Europe.
We never pass off an estimate as a benchmark. Commissions never change which hardware we recommend. How we calculate
Price details
$1,395€1,246
How we priced this
Used price, the typical price of second-hand cards that actually sold.
Source: pcprice.watch, checked 24 Oct 2026.
New price, the lowest in stock in the EU: $2,052€2,218 at GPU PRIX, 16 Oct 2026. A Zotac Trinity OC from France.
Used median €1,246 (typical €1,136–€1,334) from 235 cards sold on eBay in the past 90 days, recorded 24 Oct 2026. pcprice.watch refreshed that page on 23 Oct 2026. Not an EU shop price.
Prices excl. shipping unless marked. Euro prices incl. VAT where the seller charges it. Dollar prices are converted from our euro prices without the 21% EU VAT, before US sales tax.
A complete machine that runs it at least as fast as you read: Mac Studio M4 Max 36GB, about $2,116€1,890 (used). Comfortable~11.6 tok/s
See our picks for Qwen3.8-27B in Find hardware. Add your budget to see cheaper machines.
For comparison, an RX 9060 XT 16GB costs $522€564, new (cheapest in-stock board on GPU PRIX's German board list). Card details
If you buy through these links we may earn a commission. It doesn't change our ranking. How shop links work
As an Amazon Associate I earn from qualifying purchases.
Als Amazon-Partner verdiene ich an qualifizierten Verkäufen.
Als Amazon-Partner verdien ik aan in aanmerking komende aankopen.
Qwen3.8-27B on other graphics cards
Runs well6
The whole model fits in the graphics memory.
Doesn't fit4
Not enough graphics memory: it needs about 17.8 GB.
- RTX 3060 12GBhas 12 GB
- RTX 4060 Ti 16GBhas 16 GB
- RTX 5080has 16 GB
- RX 9060 XT 16GBhas 16 GB
More about the model itself, what it is good at and its licence: Qwen3.8-27B.
The numbers behind this answer
A model comes in several sizes. Q4 is the standard one most people use; the bigger ones answer a little better but need more memory. Each row adds up the model file, the memory a conversation takes, and some working space, and compares the total with what is free on this card.
Swipe the table sideways to see every column.
| Quality | Conversation | Result | Speed | Model file | Conversation memory | Working space | Total needed | Free memory |
|---|---|---|---|---|---|---|---|---|
| Q4 · standard (smallest that keeps quality) | ~12 pages8,192 tokens | Doesn't fit | — | 16.5 GBmeasured file | 0.5 GB | 0.8 GB | 17.8 GB | 15.4 GB |
| Q5 · a bit sharper, a bit bigger | ~12 pages8,192 tokens | Doesn't fit | — | 19.7 GB | 0.5 GB | 0.8 GB | 21.1 GB | 15.4 GB |
| Q8 · near full quality, almost twice Q4 | ~12 pages8,192 tokens | Doesn't fit | — | 29.7 GB | 0.5 GB | 0.8 GB | 31.1 GB | 15.4 GB |
| FP16 · full quality, over three times Q4 | ~12 pages8,192 tokens | Doesn't fit | — | 55.6 GB | 0.5 GB | 0.8 GB | 56.9 GB | 15.4 GB |
| Q4 · standard, longer conversation | ~49 pages32,768 tokens | Doesn't fit | — | 16.5 GBmeasured file | 2.1 GB | 0.8 GB | 19.4 GB | 15.4 GB |
Sources and method
Memory and speed on this page are our estimate, not a measured test. Checked . unsloth UD-Q4_K_M GGUF.
Dense vision-language model, Apache-2.0 (the LICENSE file is Apache 2.0, and the card metadata says apache-2.0). The card's parameter headline is 27B. Hugging Face safetensors metadata lists 27,781,427,952 parameters, all BF16. 27.78B is that count to two decimals, so our parameter count is an estimate. 64 layers in text_config: 48 linear_attention (Gated DeltaNet) and 16 full_attention (Gated Attention). 24 query heads, 4 KV heads, head dim 256. Config max_position_embeddings is 262,144. The card says that window is extensible to 1,000,000 tokens with a RoPE change; we use 262,144. Linear-attention layers are stored with a sliding window of 0 so our estimate gives them no conversation memory. The conversation-memory figure uses only the 16 full-attention layers and is an FP16 estimate. The card's citation is August 2026. Hugging Face lists the repo as created on 5 August 2026. That createdAt is repo metadata, not a dated announcement in the citation. Vision projector and the MTP drafter are separate files, not rows in the memory table. On unsloth/Qwen3.8-27B-GGUF, checked 24 Oct 2026 (https://huggingface.co/unsloth/Qwen3.8-27B-GGUF): mmproj-F16 is 0.93 GB (927,607,488 bytes) and mmproj-BF16 is 0.93 GB (931,146,432 bytes). MTP/mtp-Qwen3.8-27B-Q4_0.gguf is 1.37 GB (1,369,590,656 bytes). Main GGUF sizes are decimal GB from that tree, single files: UD-IQ2_XXS 7.27, UD-IQ3_XXS 10.93, UD-IQ4_XS 14.25, UD-Q4_K_M 16.46, UD-Q4_K_XL 17.56, UD-Q5_K_XL 20.88, UD-Q6_K 21.98, Q8_0 29.05. A separate community speed table for one RTX 3090 is stored on this row. It is not the calculator estimate.
Measured by others
Someone else ran Qwen3.8-27B and published the speed (community measurement, not ours). They may have used a smaller or bigger version than the standard size above, so their result can differ from our answer: "How they measured" has the details.
- RTX 3090, llama.cpp out of the box
- RTX 3090, llama.cpp tuned (FA + q8 KV + MTP)
- RTX 3090, stock vLLM + MTP k=4
- RTX 3090, patched vLLM + MTP k=4
- RTX 3090, patched vLLM + DFlash2 k=7
How they measured
Canonical single-GPU figures on that results page: mean of five runs, one RTX 3090, the other replica stopped, the ~1K-token prompt column, 512-token greedy completions, reasoning disabled, 250 W power cap. The page says every rung is the same 4-bit checkpoint. llama.cpp out of the box 27.6 tokens per second; llama.cpp tuned (FA + q8 KV + MTP) 42.6; stock vLLM + MTP k=4 85.6; patched vLLM + MTP k=4 167.8; patched vLLM + DFlash2 k=7 212.1. The page dates the single-replica correction 31 August. results.html was last committed on 18 September 2026, which is the published date stored here. Not a localLLMrig measurement.
Measured by others
Someone else ran Qwen3.8-27B and published the speed (community measurement, single post, not run by us). They may have used a smaller or bigger version than the standard size above, so their result can differ from our answer: "How they measured" has the details.
- 1× Intel Arc Pro B70 32GB, vLLM (cyspiegel/vllm-xpu-b70), Int4 AutoRound decode [self-reported]
- 1× Intel Arc Pro B70 32GB, llama.cpp Vulkan (Mesa 26.2.5), Q4_K_M decode [self-reported]
- 2× Intel Arc Pro B70 32GB split across B850 PCIe 4.0 x4 chipset slot, decode [self-reported]
How they measured
Self-reported by Josh Gold (github.com/killercatfish, 24 Oct 2026, quoted by name with written permission; SYCL prefill bug report in llama.cpp #30892 dated 20 Oct 2026). Headless Ubuntu 26.04 (kernel 7.0), Ryzen 5 7600X, 32 GB DDR5, B850 motherboard, two Intel Arc Pro B70 32 GB cards, one model per card. Card A runs vLLM (cyspiegel/vllm-xpu-b70) with Qwen3.8-27B Int4 (AutoRound) at about 30 tok/s decode; a 10K-token prompt takes about 7 s cold and 0.6 s warm with the prompt cache. Card B, in the B850 chipset PCIe 4.0 x4 slot, runs llama.cpp on Vulkan (Mesa 26.2.5) behind llama-swap with Qwen3.8-27B Q4_K_M at about 15 tok/s decode and 375–500 tok/s prompt reading (Vulkan: 0 faults). One card also faulted under Intel's compute stack; a slot swap later showed that card was defective, not the slot. On 26 Oct 2026 the faulty card failed 3 of 3 in the CPU x16 slot, and the healthy card ran 13 of 13 clean in the B850 chipset PCIe 4.0 x4 slot. The defective card is going back to ASRock for replacement. A single-card llama.cpp SYCL test build (b11755, GGML_SYCL_F16=OFF) measured about 270 tok/s prompt reading verified correct to 14K tokens; GGML_SYCL_F16=ON raised prompt speed to about 640 tok/s but produced !!!! NaN garbage (llama.cpp #30892; others point at PR #30604, which he had not retested). Splitting one model across both B70 cards over the B850 PCIe 4.0 x4 chipset slot dropped decode to 5.7 tok/s. Not a localLLMrig measurement.
Questions
- Can I run Qwen3.8-27B on an RX 9060 XT 16GB?
- No: it needs about 18 GB and this card has 16 GB of graphics memory.
- How much memory does Qwen3.8-27B need?
- About 17.8 GB at the standard quality (Q4) for a conversation of up to about 12 pages of text (8,192 tokens), by our estimate: 16.5 GB for the model file (measured), 0.5 GB to hold the conversation and 0.8 GB of working space.
- How fast is Qwen3.8-27B on an RX 9060 XT 16GB?
- There is no speed to give: it doesn't fit on this card.
Speeds and memory are our estimate from the hardware specs, not a measured test. Prices are dated and may change. How we calculate