Models / Qwen3.8 Flash-Next / VRAM
How much VRAM do you need for Qwen3.8 Flash-Next?
Last updated:
Weight math from the model card, plus every measured RigBase result for Qwen3.8 Flash-Next. Leave extra room for context (KV cache).
Weight size by quant
| Quant | Bits/weight | Weights only |
|---|---|---|
| Q8_0 | 8.50 | 191.3 GB |
| Q6_K | 6.56 | 147.7 GB |
| Q5_K_M | 5.70 | 128.3 GB |
| Q4_K_M | 4.89 | 110.1 GB |
| Q3_K_M | 4.00 | 89.9 GB |
| Q2_K | 3.16 | 71.1 GB |
Bits-per-weight for GGUF rows: llama.cpp quantize README. Parameter count: Official Hugging Face model card: 125B backbone with 6B activated parameters, pl. We advise staying at Q4 or above for chat quality — see Is Q4 good enough?.
Measured speeds for Qwen3.8 Flash-Next
| Hardware | Quant | Engine | tok/s |
|---|---|---|---|
| RTX Pro 6000 Blackwell Workstation (96 GB) + RTX 3090 Ti (EV · 120 GB | IQ3_S + model draft head, all experts on the RTX P | Strata engine 0.1.39, RTX Pro 6000 | 205.263 ⚡ |
| 2x RTX 5070 Ti 16 GB · 32 GB | GSQ-RCO (64 GB edition) | capy-flash v0.7.0, 2 GPUs, agentic decod | 182 |
| RTX 5090 · 32 GB | IQ3_S (~55 GB), experts offloaded to RAM, speculat | Strata | 159.2 |
| 2x RTX 5070 Ti 16 GB · 32 GB | NVFP4 | capy-flash v0.6.0, 2 GPUs, agentic decod | 155 |
| 4x RTX 3090 · 96 GB | W4A16 + draft | vLLM TP4, 1 stream | 145.8 ⚡ |
| 2x RTX 5080 16 GB · 32 GB | NVFP4 | capy-flash v0.7.2, --expert-cpu-threads | 144.4 |
| 2x RTX 5080 16 GB · 32 GB | NVFP4 | capy-flash v0.7.2 (CUDA 13.4), agentic d | 141.3 |
| 2x Tesla V100 32 GB (PG500-216, sm_70) · 64 GB | IQ3_S | bnk v0.3.1, decode at 4K context | 121 |
| 2x Tesla V100 32 GB (PG500-216, sm_70) · 64 GB | FP4-Local | bnk v0.3.1, decode at 4K context | 114 |
| 4x RTX 3090 · 96 GB | W4A16 + draft | vLLM TP4, 1 stream | 109.5 ⚡ |
| RTX Pro 6000 Blackwell Workstation (96 GB) + RTX 3090 Ti (EV · 120 GB | IQ3_S, all experts on the RTX Pro 6000, no draft h | llama.cpp (3cf03257), RTX Pro 6000 | 102.889 ⚡ |
| AMD Radeon AI Pro R9700 32 GB (OCuLink x4) + Ryzen AI Max+ 3 · 160 GB | custom int4 experts / 6-bit trunk | strixpro (MTP depth 3 + prompt lookup) | 102 ⚡ |
⚡ = speculative decoding (MTP, DFlash…). Full list: best home rig for Qwen3.8 Flash-Next.
Related
All Qwen3.8 Flash-Next builds VRAM + KV calculator By VRAM size GGUF vs AWQ vs FP8 vs NVFP4 How much system RAM?
FAQ
How much VRAM do you need for Qwen3.8 Flash-Next?
Weights alone are about 110.1 GB at Q4_K_M (180B × 4.89 bits ÷ 8). Add memory for the KV cache (context) and the runtime. The smallest-memory measured result in RigBase is 12 GB (RTX 5070 12GB) at 11.5 tok/s. Going below Q4 usually hurts quality for chat; see the Q4 guide.
What is the fastest measured speed for Qwen3.8 Flash-Next?
2x RTX 5070 Ti 16 GB: 182 tok/s plain decode (GSQ-RCO (64 GB edition), capy-flash v0.7.0, 2 GPUs, agentic decode).
Does system RAM replace VRAM?
Only for overflow. When the whole model fits in GPU or unified memory, system RAM barely matters for decode speed. See the system-RAM guide.
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.