Alibaba, China · Released August 24, 2026
Qwen3.8-Flash-Next has 180.0B parameters (6B active per token). In BF16 its weights alone take 330 GiB; quantized to 4-bit, about 105 GiB. Each 32K-token request adds 0.77 GiB of KV cache. Cheapest way to run it today: 5 × GeForce RTX 3090 (24 GB) at $0.715 per hour.
Size your own setup in the calculator →
Alibaba, China · Released August 24, 2026ollama run qwen3.8-flash-next 181K pulls (opens in a new tab)Weights plus the KV cache of one request (BF16 cache) and the model's fixed recurrent state, in GiB. Add 1 to 3 GiB for the inference engine itself.
| Weights format | Weights only | + 8K context | + 32K context | + 128K context | + 256K context |
|---|---|---|---|---|---|
| BF16 / FP16 | 330 | 331 | 331 | 334 | 337 |
| FP8 / INT8 | 166 | 167 | 167 | 170 | 173 |
| 4-bit (GGUF Q4_K_M) | 105 | 105 | 106 | 108 | 111 |
Exact size of each quantization in unsloth/Qwen3.8-Flash-Next-GGUF (opens in a new tab), and the smallest GPU that runs it with llama.cpp, Ollama or LM Studio and an 8K-token context.
| Quantization | File size | Smallest GPU |
|---|---|---|
| BF16 | 330 GiB | More than 96 GB: see the calculator |
| Q8_0 | 175 GiB | More than 96 GB: see the calculator |
| UD-Q6_K_XL | 158 GiB | More than 96 GB: see the calculator |
| UD-Q5_K_XL | 147 GiB | More than 96 GB: see the calculator |
| UD-Q4_K_XL | 104 GiB | More than 96 GB: see the calculator |
| UD-IQ4_XS | 87.2 GiB | RTX PRO 6000 Blackwell (96 GB) |
| UD-Q3_K_XL | 83.8 GiB | RTX PRO 6000 Blackwell (96 GB) |
| UD-IQ3_XXS | 76.3 GiB | RTX PRO 6000 Blackwell (96 GB) |
| UD-Q2_K_XL | 73.5 GiB | RTX PRO 6000 Blackwell (96 GB) |
| UD-IQ1_M | 69.4 GiB | RTX PRO 6000 Blackwell (96 GB) |
| UD-IQ1_S | 67.6 GiB | RTX PRO 6000 Blackwell (96 GB) |
Read every day from Hugging Face. Split files are added together; vision projectors are left out.
Fewest GPUs that fit the weights plus one 8K-token request. Data-center GPUs use vLLM (90% of memory usable), the others llama.cpp. Several GPUs split the model with tensor or pipeline parallelism.
| GPU | BF16 / FP16 | FP8 / INT8 | 4-bit (GGUF Q4_K_M) |
|---|---|---|---|
| GeForce RTX 4090 24 GB | 16 (several servers) | 8 | 5 |
| GeForce RTX 5090 32 GB | 16 (several servers) | 6 | 4 |
| RTX PRO 6000 Blackwell 96 GB | 4 | 2 | 2 |
| Mac M5 Max 128 GB 96 GB usable | 4 | 2 | 2 |
| NVIDIA H100 80 GB | 5 | 3 | 2 |
| NVIDIA H200 141 GB | 3 | 2 | 1 |
| NVIDIA B200 (HGX) 180 GB | 3 | 2 | 1 |
For each GPU: the fewest cards that fit Qwen3.8-Flash-Next with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-10-10).
On-demand prices from RunPod, Vast.ai, Verda and Azure, checked every hour. Serving many users needs more KV cache, so more memory: size it in the calculator. Some links are affiliate links: StudioTV may earn a commission, at no extra cost to you.
Architecture: MoE 512 experts · 12 Qwen Sparse Attention layers (GQA, 2 KV × 256, + indexer keys per 4-token block) · 36 Gated DeltaNet layers · 51B n-gram embeddings. Its KV cache is 3.9× smaller than if every layer used full attention, because it keeps only a short window of tokens at full detail and compresses the rest of the sequence. Only 6B parameters are active per token, but all 180B must sit in memory: VRAM depends on the total, speed on the active part.
| Context per request | BF16 cache | FP8 cache |
|---|---|---|
| 8K tokens | 0.19 GiB | 0.10 GiB |
| 32K tokens | 0.77 GiB | 0.40 GiB |
| 128K tokens | 3.09 GiB | 1.59 GiB |
| 256K tokens | 6.19 GiB | 3.19 GiB |
Plus 110 MiB of recurrent state per request, whatever the context.
In BF16 the weights alone take 330 GiB (355 GB). In FP8 that is 166 GiB, and about 105 GiB with 4-bit quantization (Q4_K_M). Each request then adds KV cache: 0.19 GiB at 8K tokens and 0.77 GiB at 32K (BF16 cache).
No. Even in 4-bit it needs 5 RTX 4090s. In 4-bit it fits on a single NVIDIA H200 (141 GB).
On 2026-10-10, the cheapest on-demand setup is 5 × GeForce RTX 3090 (24 GB) with 4-bit (GGUF Q4_K_M) weights on Vast.ai, at $0.715 per hour (about $521.95 per month). Next: 3 × RTX 6000 Ada (48 GB) with 4-bit (GGUF Q4_K_M) weights on Vast.ai, at $1.48 per hour. Sized for one 8K-token request; prices are checked every hour.
With one request and an 8K-token context: 5 in BF16, 3 in FP8 and 2 in 4-bit (vLLM, 90% of memory usable). Serving many users at once needs more memory for their KV caches: the calculator sizes that for you.
For a model card or a README: the 4-bit size and the smallest GPU it fits on, linked to this page. Data as JSON: /api/models/qwen3-8-flash-next.json.
[](https://studiotvai.com/vram-requirements/qwen3-8-flash-next)Same family (hybrid linear attention): Qwen3.5 4B · Qwen3.5 9B · Qwen3.8 27B · Qwen3.6 35B-A3B · Qwen3.5 122B-A10B · LFM2.5 8B-A1B · LFM2 24B-A2B · Ornith 1.5 35B-A3B · Ornith 1.5 9B · Clef · Clef Flash · D1 3B · JEV 27B VL · LightOnOCR-3 4B · Agnes-3.0-Qwen
All models: Llama 3.1 8B · Mistral Small 3.2 24B · Qwen3 32B · Llama 3.3 70B · Llama 3.1 405B · Qwen3.5 4B · Qwen3.5 9B · Qwen3.8 27B · Qwen3.6 35B-A3B · Qwen3.5 122B-A10B · LFM2.5 8B-A1B · LFM2 24B-A2B · Gemma 4 12B · Gemma 4 31B · Gemma 4 26B-A4B · gpt-oss 20B · gpt-oss 120B · Llama 4 Scout 17B-16E · Llama 4 Maverick 17B-128E · Kolibri-1 · Nemotron 3 Nano 30B-A3B · Nemotron 3 Super 120B-A12B · MiniMax M2.7 · DeepSeek R1 · Kimi K2.6 · GLM-5.3 · GLM-5.3-Flash · DeepSeek V4 Flash · DeepSeek V4 Pro · DeepSeek V4.1 Flash · MiMo-V2.6 Pro · MiMo-V2.6 Flash · Kimi K3 · Mistral Large 4 · Qwen3-Coder-Next 80B-A3B · Qwen3-Coder 30B-A3B · Qwen3-Coder 480B-A35B · DeepSeek V4 Flash Vision Exp · Qwen3.8 2.4T-A95B · Ornith 1.5 397B · Ornith 1.5 35B-A3B · Ornith 1.5 9B · Atria Dawn Preview · Intern S2 397B · Clef · Clef Flash · D1 3B · Humanizer · JEV 27B VL · Mellum2.1 12B-A2.5B · LightOnOCR-3 4B · Spark-X2.5 4B · Agnes-3.0-Qwen
Estimates, not guarantees: computed from the official config.json with the same engine as the LLM VRAM Calculator. Real usage depends on your engine version and settings.