Qwen3.8-Flash-Next GPU requirements
Compare memory requirements and multi-GPU serving options for Qwen3.8-Flash-Next. Choose Fluence GPU capacity to run the model with your own runtime, application settings and access controls.
Deployment at a glance
FP8 inference, CPU offload
H100 × 4
·
80 GB
N-gram embedding table on the host
FP8 inference, expert parallelism
H200 × 8
·
141 GB
Tensor and expert parallel serving
Built for coding and multimodal tasks
Qwen3.8-Flash-Next combines sparse expert routing, hybrid attention and a large n-gram embedding table. It accepts text, images and video, with text output for coding, tool use and knowledge-work applications.
Model detail
Qwen3.8-Flash-Next
Publisher
Qwen, Alibaba Group
Main model
125B parameters; approximately 6B activated per token
Additional components
51B n-gram embeddings and 4B multi-token prediction module
Full stored checkpoint
Approximately 180B parameters
Native context
262,144 tokens
Extended context
Up to 1,000,000 with additional configuration
Official checkpoints
BF16 and FP8
License
Qwen Community License 1.0
Built for coding and multimodal tasks
Qwen3.8-Flash-Next combines sparse expert routing, hybrid attention and a large n-gram embedding table. It accepts text, images and video, with text output for coding, tool use and knowledge-work applications.
Publisher
Qwen, Alibaba Group
Main model
125B parameters; approximately 6B activated per token
Additional components
51B n-gram embeddings and 4B multi-token prediction module
Full stored checkpoint
Approximately 180B parameters
Native context
262,144 tokens
Extended context
Up to 1,000,000 with additional configuration
Official checkpoints
BF16 and FP8
License
Qwen Community License 1.0
Best used for
Best used for
Qwen3.8-Flash-Next VRAM requirements
Plan memory for the full checkpoint and runtime. The active parameter count describes computation per token; the larger stored model still affects deployment capacity.
Published weight footprint
Start with the checkpoint
Weights only. Runtime and requests need additional memory.
BF16
335 GiB
Official FP8
173 GiB
Note: These approximate weights-only figures exclude runtime state, request memory and other allocations. GPU-resident memory also depends on offload and parallelism.
GPU configurations for Qwen inference
Start with the published FP8 serving recipes, then check matching Fluence capacity. Confirm the hardware topology, host resources and runtime settings for the complete deployment.
FP8 inference
H100 × 4
·
80 GB
Serve with n-gram embedding CPU offload. Confirm sufficient host RAM and the recipe’s runtime settings.
CHECKPOINT
Official FP8
KEY REQUIREMENT
CPU offload
RUNTIME REFERENCE
vLLM
FP8 inference
H200 × 8
·
141 GB
Serve with tensor parallelism and expert parallelism. Confirm GPU topology and the recipe’s runtime settings.
CHECKPOINT
Official FP8
KEY REQUIREMENT
Expert parallelism
RUNTIME REFERENCE
vLLM
GPU configurations for Qwen inference
Start with the published FP8 serving recipes, then check matching Fluence capacity. Confirm the hardware topology, host resources and runtime settings for the complete deployment.
FP8 inference
H100 × 4
·
80 GB
Serve with n-gram embedding CPU offload. Confirm sufficient host RAM and the recipe’s runtime settings.
CHECKPOINT
Official FP8
KEY REQUIREMENT
CPU offload
RUNTIME REFERENCE
vLLM
FP8 inference
H200 × 8
·
141 GB
Serve with tensor parallelism and expert parallelism. Confirm GPU topology and the recipe’s runtime settings.
CHECKPOINT
Official FP8
KEY REQUIREMENT
Expert parallelism
RUNTIME REFERENCE
vLLM
Find matching Fluence capacity
Compare current offers by configuration, location and total rental rate. Choose a connected multi-GPU environment that meets the recipe’s requirements.
GPU configuration
Region
Price per hour
Availability
Request a custom GPU cluster
Need a different GPU setup or configuration? Send us your requirements and our team will get back to you.
Request a custom GPU cluster
Need a different GPU setup or configuration? Send us your requirements and our team will get back to you.
Plan fine-tuning and training separately
Adapting this model needs a training recipe and its own resource budget. Published ms-SWIFT guidance covers Megatron LoRA fine-tuning and reinforcement-learning post-training.
Plan fine-tuning and training separately
Adapting this model needs a training recipe and its own resource budget. Published ms-SWIFT guidance covers Megatron LoRA fine-tuning and reinforcement-learning post-training.
Estimate AI workload costs on Fluence
Compare estimated compute costs for inference, fine‑tuning and training. Choose your workload, model and GPU to explore costs across providers.
Fluence$0.42/1M tokensLowest demo cost
CoreWeave$1.26/1M tokens
AWS$1.58/1M tokens
Google Cloud$2.17/1M tokensEstimate AI workload costs on Fluence
Compare estimated compute costs for inference, fine‑tuning and training. Choose your workload, model and GPU to explore costs across providers.
Fluence$0.42/1M tokensLowest demo costCoreWeave$1.26/1M tokensAWS$1.58/1M tokensGoogle Cloud$2.17/1M tokensRun Qwen with your own serving stack
Select GPU capacity on Fluence, load the checkpoint and configure a compatible runtime. You control model serving and how your application connects to it.
1
Prepare your environment
Check GPU connectivity, host RAM and storage. Review model access and licensing, then select a runtime build that supports this checkpoint.
2
Configure model serving
Follow the model-specific serving recipe. Match checkpoint precision, parallelism and offload settings to the selected hardware, then configure reasoning and tool-call handling.
3
Verify and connect
Send a first request to your server, then connect your application securely. Test representative inputs and monitor memory, latency and throughput before increasing demand.
Run Qwen with your own serving stack
Select GPU capacity on Fluence, load the checkpoint and configure a compatible runtime. You control model serving and how your application connects to it.
1
Prepare your environment
Check GPU connectivity, host RAM and storage. Review model access and licensing, then select a runtime build that supports this checkpoint.
2
Configure model serving
Follow the model-specific serving recipe. Match checkpoint precision, parallelism and offload settings to the selected hardware, then configure reasoning and tool-call handling.
3
Verify and connect
Send a first request to your server, then connect your application securely. Test representative inputs and monitor memory, latency and throughput before increasing demand.
Why run Qwen on Fluence Cloud
Why run Qwen on Fluence Cloud
Top-tier data centers with enterprise-grade hardware
Our high-performance servers are deployed across highly available Tier III and Tier IV facilities, providing GDPR-compliant infrastructure that satisfies SOC 2 and ISO 27001 standards.
Top-tier data centers with enterprise-grade hardware
Our high-performance servers are deployed across highly available Tier III and Tier IV facilities, providing GDPR-compliant infrastructure that satisfies SOC 2 and ISO 27001 standards.
FAQ
How much VRAM does Qwen3.8-Flash-Next need?
Does the 6B active parameter count make this a small model?
Can I run it on a single 80 GB GPU?
Can I use the full context window with these configurations?
Can I fine-tune it on Fluence?
Show more
FAQ
How much VRAM does Qwen3.8-Flash-Next need?
Does the 6B active parameter count make this a small model?
Can I run it on a single 80 GB GPU?
Can I use the full context window with these configurations?
Can I fine-tune it on Fluence?
Show more
Choose GPUs for your Qwen workload
Find capacity for your selected checkpoint and deployment plan. Compare current offers or request a multi-GPU configuration for your serving or training requirements.
Choose GPUs for your Qwen workload
Find capacity for your selected checkpoint and deployment plan. Compare current offers or request a multi-GPU configuration for your serving or training requirements.