High-memory GPUs on OneInfer
Qwen3.8-Flash-Next fits a single B300 / B200 SXM at FP8. These are current OneInfer listings by VRAM, for scoping a self-host deployment.
VRAM by precision
| Precision | Approximate VRAM | Notes |
|---|---|---|
| FP8 | 172.78 GiB | 125B params × ~1 byte + ~1.38 bytes for the embedding/MTP head. Fits a single B200 SXM (192 GB) or B300 (262 GB) with headroom for KV cache. Source: MarkTechPost checkpoint report, 26 Aug 2026 — re-verify against the Hugging Face file listing before publishing. |
| BF16 | 335.28 GiB | 125B params × 2 bytes + ~10.28 GiB for the embedding/MTP head. Needs at least two B200/B300-class cards in tensor parallel. Source: MarkTechPost checkpoint report, 26 Aug 2026. |
| INT4 | ~75 GiB | 125B params × 0.5 bytes after INT4 quantization. Fits a single H100 (80 GB) with KV-cache headroom on shorter contexts; expect quality caveats on long-context reasoning. |
How these figures are calculated
Bytes-per-parameter is applied to the 125B total parameter count (6B active per token). Numbers are the parameter-only weight footprint before KV cache, activations, or runtime overhead. Multiply by ~1.2 to budget for sustained inference at the 262K native context.
Ready to test the workflow?
Create account & add creditsWeights are publicly available
Qwen3.8-Flash weights are released under the qwen-community-1.0 license and can be pulled from the Hugging Face repository linked in Sources. Confirm the license tag against your deployment context before self-hosting at scale.
Architecture
125B-parameter MoE (6B activated per token), native 262,144-token context, extensible to ~1M via YaRN. Verify the architecture detail against the Hugging Face model card before publishing; the numbers above are derived from the checkpoint file listing, not a paper.
Frequently asked questions
What GPU do I need to run Qwen3.8-Flash?
FP8 weights (172.78 GiB) fit on a single B200 SXM (192 GB) or B300 (262 GB). BF16 needs at least two B200/B300-class cards in tensor parallel. INT4 quantization fits a single H100 (80 GB) with KV-cache headroom on shorter contexts.
Is Qwen3.8-Flash open source?
Yes — the Flash weights are released under the qwen-community-1.0 license. Confirm the license tag and any use restrictions against the official Hugging Face repository before self-hosting at scale.
How is VRAM estimated?
Bytes-per-parameter is multiplied by the 125B total parameter count (6B active per token). The figures are the parameter-only weight footprint before KV cache, activations, or runtime overhead — multiply by ~1.2 for sustained 262K-context inference.
Where is the live Qwen3.8-Flash model page?
The canonical model page with current OneInfer pricing, capabilities, and availability is /models/Qwen/Qwen3.8-Flash. This page is a focused facet of that entity, not a replacement for it.
How should I treat benchmark or price claims?
Check each claim's provenance label and observed date. Vendor-reported and independently verified numbers are shown as separate evidence classes on this hub.
Put Qwen3.8-Flash to work
Fund a controlled evaluation, start with a prepared prompt, and measure quality and cost on your own workload.