Qwen 4: release date, models and benchmarks

Qwen 4 isn't out yet, but you can already run its architecture. Qwen3.8-Flash-Next is Qwen's preview of it, and it's ready here for your hardware. On release day, the Qwen 4 GGUF files land here too.

By Stealth Models.

Run Qwen on your hardware

Qwen3.8-Flash-Next: the Qwen 4 preview

Qwen3.8-Flash-Next already uses the new design: a learned n-gram memory next to sparse attention, with 512 experts of which 10 work on each token. It reads 262,144 tokens natively and up to 1 million with YaRN.

It fits on far less than its size: only 4 to 5.5 GB has to live on the GPU. The experts wait in system RAM, and the n-gram table (29 GB, 54 GB in Q5) stays on your SSD. Use a current llama.cpp build: b11971 from 28 October or newer.

Pick your settings

All rows use a 4-bit KV cache (-ctk q4_0 -ctv q4_0): the whole cache is only about 2 GB at 256k and 8 GB at 1M.

--n-cpu-moe is how many of the 48 expert layers stay in system RAM. Out of memory? Raise it by 2. VRAM to spare? Lower it by 1.

Speed and memory options

• n-gram table on SSD or in RAM. -ot "^per_layer_token_embd\.weight$=CPU" --lazy-mode on keeps it on your SSD: no VRAM, no RAM, a few small reads per token, so use a fast NVMe drive. With RAM to spare (29 GB, 54 GB for Q5), --lazy-mode off keeps it in memory for slightly faster lookups.
• ngram-mod, free speed. --spec-type ngram-mod drafts tokens from text already in your context. It costs no VRAM or RAM and helps most when the answer repeats your input: code edits, summaries, long documents.
• MTP, the model's own draft head. Qwen trained one multi-token prediction layer, and Unsloth measures 1.3 to 1.7 times faster output with it. It takes 2.8 GB of VRAM (1.9 GB with the Q4_K_M file), so raise --n-cpu-moe by 3 on Q2 and by 2 on Q4 and Q5. It runs today on Unsloth's llama.cpp branch qwen4exp/mtp; support in official llama.cpp is still in review. Add -md qwen4/Qwen3.8-Flash-Next/MTP/mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf --spec-type draft-mtp --spec-draft-n-max 5, or combine both with --spec-type draft-mtp,ngram-mod.
• 1M context. Add --rope-scaling yarn --rope-scale 4 --yarn-orig-ctx 262144 -c 1000000 -ub 512. For 512k use --rope-scale 2 -c 524288. Qwen advises YaRN only when you need the length.

Download and run

hf download stealth-model/Qwen3.8-Flash-Next__Qwen4-news --local-dir qwen4 \
  --include "Qwen3.8-Flash-Next/Q2_K_XL/*" "Qwen3.8-Flash-Next/mmproj-Q8_0.gguf"

GPU plus system RAM (here a 24 GB GPU with 64 GB of RAM, 256k context):

llama-server \
  -m qwen4/Qwen3.8-Flash-Next/Q2_K_XL/Qwen3.8-Flash-Next-Q2_K_XL-00001-of-00003.gguf \
  --mmproj qwen4/Qwen3.8-Flash-Next/mmproj-Q8_0.gguf \
  -ngl all --n-cpu-moe 34 \
  -ot "^per_layer_token_embd\.weight$=CPU" --lazy-mode on \
  -fa on -ctk q4_0 -ctv q4_0 -c 262144 \
  -ub 2048 -b 2048 -t 8 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0

• -t is your number of physical cores; fast RAM matters most here.
• On a 16 GB GPU, use -ub 1024.

AMD Strix Halo or NVIDIA DGX Spark (128 GB unified, 256k context):

llama-server \
  -m qwen4/Qwen3.8-Flash-Next/Q4_K_XL/Qwen3.8-Flash-Next-Q4_K_XL-00001-of-00004.gguf \
  --mmproj qwen4/Qwen3.8-Flash-Next/mmproj-Q8_0.gguf \
  -ngl all -ot "^per_layer_token_embd\.weight$=CPU" --lazy-mode on \
  -fa on -ctk q4_0 -ctv q4_0 -c 262144 -ub 2048 -b 2048 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0

Q4_K_XL takes about 89 GB this way, and fits 1M context too (about 96 GB). Q5_K_XL fits at 256k if your system gives the GPU about 110 GB. With 192 GB or more, swap --lazy-mode on for --lazy-mode off.

Qwen's sampling: thinking --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0, instruct --temp 0.7 --top-p 0.8 --top-k 20 --presence-penalty 1.5.

What's your setup? Tell us your GPU, RAM and speed in the Community tab, and we'll tune the table.

---

See where every model stands on our StealthMark leaderboard.

When will Qwen 4 be released?

No date yet. Alibaba confirmed on 22 September that Qwen 4 is in training, and by 23 October hardware teams were already preparing kernels for Qwen4-27B and Qwen4.0-Max. All eyes are on Apsara Europe in Amsterdam on 29 October.

Which Qwen 4 models are coming?

Four names were on Alibaba's screen: Qwen4-Max, Qwen4-Plus, Qwen4-Flash and Qwen4-27B, with 27B first in line for open weights. A small 35B-A3B looks likely too.

Will Qwen4-27B run on my PC?

Kernel code calls it dense, built for one or two GPUs at 4-bit. Its predecessor, Qwen3.8-27B, is the best local model in our StealthMark tests today. Compare it with four other open models

Is Kalmia Qwen 4?

Maybe. Kalmia is an anonymous model on Arena that trips over the same tokens Qwen models do. That points to the Qwen family, not yet to Qwen 4 itself. See the tokenizer evidence

Are the leaked Qwen 4 benchmarks real?

None of them traces back to a real source. One chart never appeared on Artificial Analysis, and one number comes from a paper about Qwen3.8-27B.

How good will Qwen 4 be?

Our forecast: about ten points above each Qwen 3.8 model. That puts Qwen4-27B ahead of today's best open model and Qwen4-Max within reach of GPT-6 Astra. Our benchmark forecast, model by model

---

The full Qwen 4 guide · StealthMark leaderboard · Follow us for the release
Text and images: CC BY 4.0, stealthmodels.com. GGUF files: Qwen Community License 1.0.