What GPU do you need to run Qwen3.8-Flash-Next?
A new Framework Desktop ($6,799). No priced GPU card holds Qwen3.8-Flash-Next (about 121.1 GB at Q4); this system holds it with room for its full 256K context, at about 42 tok/s estimate. 2 of the 157 GPUs and systems we track fit it.
180B parameters, 6B active (mixture-of-experts) (Hugging Face, accessed 2026-10-16). Q4_K_M GGUF: 119.6 GB (source). Architecture config (Hugging Face config.json): 12 attention layers (the rest are recurrent) × 2 KV heads × 256 head dim, so about 24 KB of FP16 KV cache per token of context (how we estimate context).
Model card: 125B parameters with 6B activated, plus 51B n-gram embedding and 4B MTP.
Other places to get it
Checked Nov 14, 2026. We list only copies we opened and checked. Check file hashes for copies from other uploaders.
- ModelScope: Lab's own account
Single GPUs that fit, cheapest first
| Single GPU that fits | VRAM | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | 192 GB | ~42 tok/s | 256K (full) | low $6,799 |
| NVIDIA DGX Spark (128GB) | 128 GB | ~48 tok/s | 256K (full) | low $6,949.99 |
Every device that runs it
Graphics cards, Macs and mini PCs from the wider hardware list behind My rig, smallest usable memory first. Not sure about yours? Check your exact rig.
Show all 10 devices that run it
| Device | Type | Memory | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|---|
| AMD Ryzen AI Max 385 (128GB) | Mini PC | 128 GB | ~40 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max 390 (128GB) | Mini PC | 128 GB | ~40 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max+ 395 (128GB) | Mini PC | 128 GB | ~40 tok/s | 256K (full) | low $3,499.99 |
| NVIDIA DGX Spark (128GB) | Mini PC | 128 GB | ~48 tok/s | 256K (full) | low $6,949.99 |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 128GB) | Laptop | 128 GB | ~52 tok/s | 256K (full) | No price yet |
| Apple M2 Ultra 192GB | Mac | 192 GB (~144 usable) | ~76 tok/s | 256K (full) | No price yet |
| Apple M3 Ultra 256GB | Mac | 256 GB (~192 usable) | ~77 tok/sllama.cpp-style estimateMLX (measured, community), with speculative decoding (MTP/DFlash): 81.2–92.4 tok/s at 1K–16K (M3 Ultra 60c) | 256K (full) | No price yet |
| Apple M5 Ultra 256GB | Mac | 256 GB (~192 usable) | ~89 tok/s | 256K (full) | No price yet |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | Mini PC | 192 GB | ~42 tok/s | 256K (full) | low $6,799 |
| Apple M5 Ultra 512GB | Mac | 512 GB (~384 usable) | ~89 tok/s | 256K (full) | No price yet |
Multi-GPU: cards that need two or more
| Cards | Total VRAM | Speed estimate | Max context estimate |
|---|---|---|---|
| 4 × NVIDIA GeForce RTX 5090 | 128 GB | ~151 tok/s | 256K (full) |
| 3 × NVIDIA RTX A6000 | 144 GB | ~100 tok/s | 256K (full) |
Measured on Macs with MLX (community)
MLX (measured, community): decode speeds people ran with MLX through oMLX and submitted to its public leaderboard. They are kept apart from this site's llama.cpp-style estimate, which is for a GGUF file in llama.cpp and is unchanged. The plain figure has no speculative decoding, at the shortest context the run used; runs with speculative decoding (MTP/DFlash) are shown separately, as a range, because their speed depends on how many drafted tokens are accepted, which varies with the prompt.
| Mac | llama.cpp-style estimate | MLX plain decode | With speculative decoding (MTP/DFlash) | Long context, one run |
|---|---|---|---|---|
| Apple M5 Ultra 96GB | doesn't fit (site rule) | — | 66.6–106.1 tok/s at 1K–16K (M5 Ultra 64c) | 1K 66.6 → 16K 106.1 tok/s; prefill 1101–3807 tok/s (M5 Ultra 64c, speculative run) |
| Apple M3 Max (40-core GPU) 128GB | doesn't fit (site rule) | — | 47.4–57.3 tok/s (runs 1, 2, 3) at 1K–32K (M3 Max 40c) | 1K 54.4 → 32K 47.4 tok/s; prefill 528–709 tok/s (M3 Max 40c, speculative run) |
| Apple M4 Max (40-core GPU) 128GB | doesn't fit (site rule) | — | 44.4–78.0 tok/s (runs 1, 2, 3) at 1K–16K (M4 Max 40c) | 1K 71.6 → 16K 77.4 tok/s; prefill 607–827 tok/s (M4 Max 40c, speculative run) |
| Apple M5 Max (40-core GPU) 128GB | doesn't fit (site rule) | — | 62.4–69.1 tok/s at 1K–4K (M5 Max 40c) | 1K 69.1 → 4K 62.4 tok/s; prefill 1053–1993 tok/s (M5 Max 40c, speculative run) |
| Apple M3 Ultra 256GB | ~77 tok/s | — | 81.2–92.4 tok/s at 1K–16K (M3 Ultra 60c) | 1K 92.4 → 16K 90.7 tok/s; prefill 913–1082 tok/s (M3 Ultra 60c, speculative run) |
Long context: decode at the shortest and longest context of one run, and that run's prompt-processing (prefill) range. Agent jobs re-read long prompts, so prefill (prompt processing) speed matters as much as decode there. Community-submitted to oMLX, accessed 2026-10-26; each figure links its run(s). MLX 4-bit files, not the GGUF the estimate uses. About the MLX data
Speed calculator
estimateEstimate = bandwidth ÷ GB read per token (the model's listed GGUF file: Q4_K_M for most models, otherwise the quant named in the list; scaled by active/total parameters for MoE), scaled by a per-platform efficiency and fixed per-token costs calibrated to 45 measured llama.cpp benchmarks (2025–26). Fit = file size + 1.5 GB headroom ≤ total VRAM (on Macs, the part of unified memory macOS lets the GPU use by default), and room for at least 4K tokens of context with an 8-bit KV cache. For a model with no KV cache data, only the file and headroom are checked, so a fit says nothing about context. Multi-GPU: VRAM adds, speed stays one card's (a layer split runs the cards in turn); max context uses the total. Speed is for short context. Max context = what is left after the file and a 1 GB runtime reserve ÷ KV cache bytes per token, default FP16 and 8-bit (q8_0) cache. How speed works · how context works