Deployment guide
Cheapest way to deploy Qwen3.5-397B-A17B in 2026
6 providers compared. API token-pricing, dedicated capacity, and rented GPU costs side-by-side, normalized to monthly cost.
Cheapest API
$0.600 / 1M input tokens
at Nebius
Cheapest GPU rental
$3.50 / hour
at Nebius on H200 SXM
Provider rate cards
6 providers compared
| Provider | Hardware | Region | Commitment | Source | |||
|---|---|---|---|---|---|---|---|
 Nebius | H200 SXM | 8 | 141 | Netherlands | $3.50 | On-demand | Source ↗ |
| Nebius | B200 SXM | 8 | 192 | Netherlands | $5.50 | On-demand | Source ↗ |
 Azure | RTX PRO 6000 | 16 | 96 | Multiple regions | $5.50 | On-demand | Source ↗ |
| Nebius | B300 SXM | 8 | 288 | Netherlands | $6.10 | On-demand | Source ↗ |
 Verda | H200 SXM | 8 | 141 | Finland | $27.12 | On-demand | Source ↗ |
| Verda | B300 SXM | 4 | 288 | Finland | $27.96 | On-demand | Source ↗ |
| Verda | B200 SXM | 8 | 192 | Finland | $39.12 | On-demand | Source ↗ |
 CoreWeave | H200 SXM | 8 | 141 | United States | $50.44 | On-demand | Source ↗ |
| Azure | MI300X | 8 | 192 | Multiple regions | $57.60 | On-demand | Source ↗ |
 AWS | H200 SXM | 8 | 141 | Multiple regions | $63.30 | On-demand | Source ↗ |
| CoreWeave | B200 SXM | 8 | 192 | United States | $68.80 | On-demand | Source ↗ |
| Azure | H200 SXM | 8 | 141 | Multiple regions | $84.80 | On-demand | Source ↗ |
Break-even chart
API vs. GPU rental
The crossover point at which renting a GPU full-time becomes cheaper than paying per token: ~40M tokens/day.
Your monthly cost
Top 5 cheapest for your workload
Adjust the assumptions below — token volume, input/output ratio, days and hours of usage — to see how the cheapest options shift.
| Rank | Provider | Pricing | Hardware | Monthly |
|---|---|---|---|---|
| #1 | Nebius | GPU · On-demand | H200 SXM | $2.52k |
| #2 | Nebius | GPU · On-demand | B200 SXM | $3.96k |
| #3 | Azure | GPU · On-demand | RTX PRO 6000 | $3.96k |
| #4 | Nebius | GPU · On-demand | B300 SXM | $4.39k |
| #5 | Nebius | API | — | $4.50k |
Spec sheet
Qwen3.5-397B-A17B at a glance
- VRAM (native precision)
- 807 GB
- Parameters
- 403.3979B
- Native precision
- bf16
- Context length
- —
- License
- apache-2.0
- Knowledge cutoff
- —
- Modalities
- text, vision
- Access type
- Open source
- EU developed
- —
- Origin country
- —
Why these numbers
A second opinion on the data
Hardware footprint
Qwen3.5-397B-A17B is a 403.3979B-parameter model that needs 807 GB VRAM at bf16 when self-hosted at native precision (needs 8x H100 or larger cluster). Quantization to int8 typically halves the VRAM requirement; int4 quarters it, at modest accuracy cost. 5 GPU rental providers in nfer's index currently offer hardware that fits this model at native precision.
Cheapest path today
For Qwen3.5-397B-A17B: The cheapest API offering is Nebius at $0.60/1M input + $3.60/1M output tokens. The cheapest GPU rental that fits the model is Nebius on H200 SXM at $3.50/hour. The break-even point between paying per token and renting a GPU depends on your daily volume — see the chart above.
Licensing and fit
Released under the apache-2.0 license, Qwen3.5-397B-A17B ships with a context length not specified; open-source weights are publicly available.
FAQ
Common questions
What's the cheapest way to host Qwen3.5-397B-A17B?
The cheapest API option for Qwen3.5-397B-A17B in nfer's index is Nebius at $0.600/1M input + $3.600/1M output tokens. For self-hosted workloads, the cheapest GPU rental that fits is Nebius on H200 SXM at $3.50/hour. The right choice depends on your daily token volume — see the break-even chart on this page.How much VRAM does Qwen3.5-397B-A17B need?
Qwen3.5-397B-A17B 807 GB at native precision; roughly 404 GB at int8 and 202 GB at int4. Native precision is bf16. Quantization roughly halves (int8) or quarters (int4) the VRAM footprint at modest accuracy cost.Can I use Qwen3.5-397B-A17B commercially?
Yes — released under the apache-2.0 license, which permits commercial use.What's the difference between API and GPU rental for Qwen3.5-397B-A17B?
Token-priced API providers (like Nebius) bill per million input/output tokens — best for low or bursty volume. Renting a GPU (e.g. Nebius at $3.50/hour) is a flat ~$2520.00/month regardless of usage — better economics once you sustain enough tokens per day to justify the fixed cost. The break-even chart on this page shows the exact crossover point.Is Qwen3.5-397B-A17B available with EU data residency?
Qwen3.5-397B-A17B is not a European-developed model. 2 EU-owned providers offer hosting in nfer's index — filter on EU sovereignty in the comparator to see them.
Prices last updated · 2026-04-30