By LLM Hardware team · Methodology ·
Checked 9 October 2026 · 111.33 GB Unsloth UD-Q4_K_XL
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Key fact: Qwen3.8-Flash-Next at Q4 is about 110 GB by our VRAM formula on 180B, and Unsloth’s UD-Q4_K_XL file was 111.33 GB. A 32 GB GPU does not fit it.
How we got it: Qwen/Qwen3.8-Flash-Next card: 125B with 6B activated, plus 51B n-gram embeddings and 4B MTP. 180 × 0.5 × 1.2 + 2 = 110. Hugging Face tree for unsloth/Qwen3.8-Flash-Next-GGUF, summed 9 October 2026: UD-Q4_K_XL 111.33 GB, UD-IQ4_XS 93.68 GB.
Checked:
Key fact: The Ollama tag qwen3.8:27b is the 27B model, not Flash-Next.
How we got it: ollama.com/library/qwen3.8 on 9 October 2026 listed qwen3.8:27b, qwen3.8:27b-mlx, and qwen3.8:latest. No Flash-Next tag was on that page.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
Qwen/Qwen3.8-Flash-Next is the open-weight file. Qwen Cloud’s Qwen3.8-Flash is the hosted API Qwen describes as based on these weights, with a 1M context and official tools. Buying a GPU does not subscribe you to that API, and calling the API does not put 111 GB of GGUF on your disk. The local path is llama.cpp or Unsloth Desktop, plus a machine that can hold the file. The VRAM calculator uses the same formula as the table below. Sibling pages: the Qwen3 size guide and the Strix Halo mini PC guide.
The model card, read 9 October 2026, lists a causal language model with a vision encoder, 48 layers, 512 experts with 10 routed plus 1 shared active, and a native context of 262,144 tokens. The parameter line is 125B with 6B activated, plus 51B of n-gram embeddings and 4B of multi-token prediction. Active parameters change speed. They do not shrink the file. The Hugging Face license field is other, and the LICENSE file in the repo is Qwen Community License 1.0. That is not Apache-2.0. Qwen3.8-27B, on its own card, is Apache-2.0 and is a different download.
| Count | Q4 formula | What it leaves out |
|---|---|---|
| 125B MoE only | 77.0 GB | Skips the 51B n-gram table. Too small for the published Q4 file. |
| 125B + 51B n-gram | 107.6 GB | Still omits the 4B MTP head. |
| 180B (125 + 51 + 4) | 110.0 GB | Short context. Matches the 111.33 GB UD-Q4_K_XL file closely. |
Formula: parameters × 0.5 × 1.2 + 2, which is parameters × 0.6 + 2 at Q4. Put 180 in the calculator when you want the full resident set. Put 125 only if you are testing the MoE body with the n-gram table offloaded, and then use a file size that matches that split. We did not find an Unsloth quant that is the 77 GB “MoE only” case.
These are shard sums from unsloth/Qwen3.8-Flash-Next-GGUF on 9 October 2026. The first shard of several quants is a few megabytes; the total is the whole set. MTP files are an optional speed head, not a substitute for the model. The F16 mmproj was 0.90 GB if you need the vision tower.
| Unsloth quant | Sum | Fits |
|---|---|---|
| UD-IQ1_S | 72.55 GB | 128 GB unified. Misses 64 GB. |
| UD-Q2_K_XL | 78.87 GB | 128 GB. Tight on a 96 GB pool once the OS is counted. |
| UD-IQ3_XXS | 81.96 GB | 128 GB. |
| UD-IQ4_XS | 93.68 GB | Comfortable file on 128 GB. Does not fit 32 GB. |
| UD-Q4_K_XL | 111.33 GB | 128 GB with a short context. Not a 24 GB or 32 GB card. |
| Q8_0 | 188.23 GB | Above 128 GB. |
| BF16 | 354.03 GB | Datacenter or a much larger host. |
An earlier note on this site put Flash-Next near 70–80 GB at Q4. That range matches UD-IQ1_S and UD-Q2_K_XL, not the Q4 file. Use 111.33 GB for UD-Q4_K_XL.
A 16 GB, 24 GB, or 32 GB discrete GPU does not run any quant in the table, including UD-IQ1_S. The RTX 5090 is 32 GB. It fits Qwen3.6-35B-A3B at Q4 (23.0 GB). It does not fit Flash-Next. This guide puts no Amazon link on a 24 GB or 32 GB card for this model.
The machine in the catalog that matches the file is the GMKtec EVO-X2, Ryzen AI Max+ 395, 128 GB LPDDR5X, about 256 GB/s, ASIN B0F53MLYQ6. UD-IQ4_XS at 93.68 GB is the quant to load there. UD-Q4_K_XL at 111.33 GB leaves roughly 16 GB for the OS, the runtime, and a short KV cache, so treat it as tight. Bandwidth at ~256 GB/s means this is a slow interactive box, not a 5090. There is no CUDA. llama.cpp with Vulkan or ROCm, or LM Studio if that build loads the architecture, is the stack. Unsloth’s card says llama.cpp or Unsloth Desktop. Check current price on Amazon.
GMKtec EVO-X2 128 GB on AmazonAffiliate link. Confirm the offer before you buy. Check current price on Amazon.
The Beelink GTR9 Pro is the same 128 GB class when the listing is actually that machine. That is a different computer. ASIN B0FPQQYWQ1 still titled as the GTR9 Pro 128 GB Max+ 395 and was currently unavailable, with no buy box. This page does not link either ASIN. A 36 GB Mac Studio fits Qwen3.8-27B at Q4 (18.2 GB). It does not fit Flash-Next. Check current price on Amazon.
Download every shard of one quant into the same folder. Unsloth’s card says llama.cpp opens the -00001-of-0000N file and finds the rest. This guide prints no Ollama pull, because the library did not have a Flash-Next tag on 9 October 2026. ollama run qwen3.8:27b is the 27B model at about 18 GB. In LM Studio, pick the same Unsloth quant only after the app’s release notes say this architecture loads. The optional MTP head Unsloth ships as mtp-Qwen3.8-Flash-Next-shared-Q4_K_M was 1.91 GB. It speeds decode. It is not the model.
llama-cli -m Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00003.gguf -ngl 99
That path is the first shard name from the Unsloth tree. Use UD-Q4_K_XL’s first shard only when the host has room past 111.33 GB.
No. Qwen’s card says Qwen3.8-Flash is the managed Qwen Cloud API built on Qwen3.8-Flash-Next, with a 1M context and official built-in tools. The open weights are Qwen/Qwen3.8-Flash-Next on Hugging Face. Downloading the weights does not give you that API.
Count the whole resident set, not the 6B that are active. The card lists 125B with 6B activated, plus 51B of n-gram embeddings and a 4B MTP head. 180 × 0.5 × 1.2 + 2 = 110 GB. On 9 October 2026 Unsloth’s UD-Q4_K_XL shards summed to 111.33 GB. A 24 GB or 32 GB GPU does not fit that file. A 128 GB unified-memory box can hold it with little room left for the OS and KV cache. UD-IQ4_XS was 93.68 GB and is the more comfortable 128 GB file.
On 9 October 2026 the Ollama library page for qwen3.8 listed qwen3.8:27b, qwen3.8:27b-mlx, and qwen3.8:latest. It did not list a Flash-Next tag. Do not run qwen3.8:27b and expect these weights. Unsloth’s GGUF card says to use llama.cpp or Unsloth Desktop, and to point the loader at the first shard of the quant you downloaded.
The GMKtec EVO-X2 128 GB (ASIN B0F53MLYQ6) has 128 GB of LPDDR5X at about 256 GB/s. UD-IQ4_XS at 93.68 GB fits with room. UD-Q4_K_XL at 111.33 GB fits the file and leaves little for macOS-class overhead or a long context. It is not CUDA. The old GTR9 ASIN B0GQXDCKN1 loaded a Beelink SER3 that day, so this guide does not link it. Check current price on Amazon.
EmbeddingGemma 2 is a separate Google embedding model, about 740M parameters, not a chat replacement for Flash-Next. The Q4 formula on 0.74B is about 2.4 GB, so it fits the machines above and also fits an 8 GB GPU. Weights: google/embeddinggemma-2.