Qwen 3.8 Flash RAM and VRAM requirements (256K context)

The short answer

The model is a 125B-total / 6B-active MoE. The smallest file we recommend is 103.7 GB, so no single Radeon holds it. The busiest layers sit in VRAM, the rest of the experts sit in system RAM, and RAM is the part you size. More RAM buys a bigger, smarter, much slower model.

What each setup needs at 256K

Sizes are Linux GiB, shown as GB. File sizes from the Hugging Face listing: Q4 103.7, Q5 147.4, Q6 157.5 GB. RAM numbers are calculator estimates: the part of the file that does not fit in VRAM, plus about 7.6 GB KV at 262,144 tokens (f16), runtime, 8 GB OS, and 16 GB headroom. Q4, Q5 and Q6 only. Strix Halo memory is soldered — a DDR5 kit will not help, and a 64 GB Halo cannot hold the Q4 file.

How much VRAM does it need?

No single card holds it: Q4 is 103.7 GB plus 7.6 GB of KV cache at 256K. VRAM decides how much of the model runs at GPU speed (an R9700 reads its memory at 640 GB/s, AMD spec). System RAM holds the rest. Two 32 GB cards are the reference setup. 128 GB on two R9700s can load Q6 with about 4 GB spare on a leaner estimate; we size Q6 at 192 GB so you are not running at the edge.

• R9700
• One vs two GPUs

Is more RAM worth it?

For most people, no. 96 GB runs Q4 at full context on two R9700s. More RAM buys a larger, higher-precision file (Q5/Q6, or a bigger model like GLM 5.3 Flash at 192 GB), but more of it lives in system RAM, so it answers more slowly (estimates: Q5 ~21–25 tok/s, Q6 ~19–23 on two sticks, vs ~27–33 for Q4). Get more only if you want the most capable model you can run at home and you are fine with it being slower. 192 and 256 GB need four sticks. Ryzen 9000 is rated for DDR5-5600 with two sticks and DDR5-3600 with four (AMD spec). On the local desktop four-stick rig, llama.cpp dropped from 18.5 to 15.9 tok/s going from DDR5-5200 to 3600 (Strata 32.8 to 32.0).

• Full offload table
• GLM 5.3 Flash requirements

Why we skip the IQ4_XS file

The unsloth listing also has a UD-IQ4_XS file at 87.2 GB, which looks like a cheaper 4-bit option for 64–96 GB machines. Its GGUF header says otherwise: 31.6 of its 55.4 GB of routed experts are IQ3_S (3-bit) and the 26.8 GB n-gram table is IQ4_NL. Most of the model is below 4-bit, so it falls under our Q4-and-up rule. That is a header read, not a quality test.

• Quants

The honest pick

Already have two 32 GB cards and 96 GB or more RAM: buy nothing. Run UD-Q4_K_XL at 262,144. One R9700 and 64 GB or less: buy nothing yet — run the 35B instead. It fits fully on one card at 256K and decodes about 100 tok/s on the local desktop. Flash on one card needs a 128 GB kit and we have not measured its speed. 16–24 GB with 128 GB already: Q4 should load with expert offload; it is slow and not measured here. Building a two-R9700 rig for Flash: size RAM at 96 GB (2×48) so two sticks keep full DDR5-5600 speed. No GPU buy button on this page.

• 96 GB Flare X5 kit
• Run the 35B instead
• Motherboards

Launch line (Q4+, 256K, two cards)

--fit places as many expert layers on the two cards as fit and keeps the rest in RAM. Faster engine on the same files: Strata.

./build-vk/bin/llama-server \
  -m /path/to/nvme/Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf \
  -c 262144 --cache-type-k f16 --cache-type-v f16 -ngl 99 --fit on \
  -t <physical-cores> --host 127.0.0.1 --port 8080

• llama.cpp Flash recipe
• Backends

Our RAM pick

G.SKILL Flare X5 96 GB (2×48 GB) DDR5-5600 CL40, F5-5600J4040D48GX2-FX5. AMD EXPO. Two sticks keep the rated DDR5-5600. We omit a street price unless we read the listing today. Paid link: we may earn a commission if you buy through it, at no extra cost to you.

Check the 96 GB kit at Newegg

FAQ

How much RAM does Qwen 3.8 Flash need?

At 256K on two 32 GB R9700s, Q4 needs about 74 GB of system RAM (estimate), so a 96 GB DDR5 kit. Q5 needs about 119 GB (128 GB kit) and Q6 about 129 GB (192 GB kit). With one 32 GB card, Q4 needs a 128 GB kit.

Flash requirements · Offload table

How much VRAM does Qwen 3.8 Flash need?

No single Radeon holds it: the Q4 file is 103.7 GB plus 7.6 GB of KV at 256K. It runs with the busiest layers in VRAM and the remaining experts in system RAM. Two 32 GB cards are the reference setup.

R9700

Can I run Qwen 3.8 Flash with 64 GB of RAM?

Not at Q4 or above with full 256K on Radeon cards up to 2×32 GB: two R9700s need about 74 GB of system RAM for Q4. Below Q4 the quality drops and we do not recommend it. With 64 GB, run Qwen3.6-35B, which fits fully on one 32 GB card.

35B requirements

Is more than 96 GB of RAM worth it for Qwen 3.8 Flash?

Usually not. 128–256 GB lets you load Q5/Q6 or a larger model, but more of it sits in system RAM, so it runs slower. It is worth it only if you want the most capable model you can run and accept the slower speed.

Flash RAM offload

Sources

• S-HF-TREE — unsloth/Qwen3.8-Flash-Next-GGUF file sizes (GiB) — third-party · checked 2026-11-02 · HF tree, UD-Q4_K_XL 103.7 / UD-Q5_K_XL 147.4 / UD-Q6_K_XL 157.5 GiB
• S-IQ4 — UD-IQ4_XS GGUF header — mostly IQ3_S experts — third-party · checked 2026-11-02 · header read 2026-10-30; 31.56 of 55.43 GiB of routed experts are IQ3_S
• S-LAB — Local desktop Flash Q6 pair — Strata vs llama.cpp — our lab · checked 2026-11-02 · Measured on the local desktop
• S-OFFLOAD — Flash RAM offload kits — our lab · checked 2026-11-02 · site calculator + measured DIMM pair