DeepSeek V4.1 Flash: Self-Hosting Estimate
DeepSeek V4.1 Flash (DeepSeek) on different GPUs: configuration, estimated performance and use cases.
Data updated: 2026-09-16On 2026-10-20 DeepSeek open-sourced DeepInfer (Apache-2.0), the inference framework behind its own production service. It currently supports only this model, on NVIDIA SM100 (B200/B300) and Ascend 950. The release covers offline inference only; an online service still needs a KV cache system, prefill/decode disaggregation and load balancing, so serving an API today still means a general engine such as vLLM or SGLang. Compare with vLLM / SGLang →
B300 288GB
288 GB · 8,000 GB/s · Top-tier Blackwell card (no official channels to China)- Suggested precision
- FP16/BF16 (full precision)
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 2,000 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 3,400 tok/sin 73,125 tok/s · many concurrent requests
- Hardware list price
- $299,058–$498,429Cloud rental (live) ×8: $63.08–$63.08/hr (2026-11-04)
- Hardware market price
- $1,047,167
- Typical use
- Enterprise-grade ultra-large-scale training and high-concurrency inference clusters
GB200 192GB (single card equivalent)
192 GB · 8,000 GB/s · Top-tier Blackwell card (no official channels to China)- Suggested precision
- INT8
- GPUs needed
- 8
- Est. decode speed
- 4,000 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 6,800 tok/sin 81,250 tok/s · many concurrent requests
- Hardware list price
- $239,544–$359,316
- Hardware market price
- $718,058–$1,436,115
- Typical use
- Ultra-large-scale trillion-parameter model training and inference clusters (networked with Grace CPU)
B200 180GB
180 GB · 8,000 GB/s · Top-tier Blackwell card (no official channels to China)- Suggested precision
- INT8
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 4,000 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 6,800 tok/sin 73,125 tok/s · many concurrent requests
- Hardware list price
- $299,058–$498,429Cloud rental (live) ×8: $59.58–$100.08/hr (2026-11-05)
- Hardware market price
- $448,786–$523,584
- Typical use
- Enterprise-grade large-scale online inference and training
H200 141GB
141 GB · 4,800 GB/s · Top-tier export-controlled card (channels to China are volatile)- Suggested precision
- INT8
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 2,400 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 4,080 tok/sin 32,142 tok/s · many concurrent requests
- Hardware list price
- $279,837–$499,708Cloud rental (live) ×8: $23.93–$36.86/hr (2026-11-05)
- Hardware market price
- $224,393–$233,369
- Typical use
- Enterprise-grade production environment high-concurrency inference
H100 SXM5 80GB
80 GB · 3,350 GB/s · Top-tier export-controlled card (no official channels to China)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 3,350 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 5,695 tok/sin 32,142 tok/s · many concurrent requests
- Hardware list price
- $239,860–$831,514Cloud rental (live) ×8: $15.47–$32.90/hr (2026-11-05)
- Hardware market price
- $320,732
- Typical use
- Enterprise-grade production environment high-concurrency inference/training
H800 SXM 80GB
80 GB · 3,350 GB/s · Top-tier export-controlled card (discontinued)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8priced as 1 × 8-GPU HGX node
- Est. decode speed
- 3,350 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 5,695 tok/sin 32,142 tok/s · many concurrent requests
- Hardware list price
- $293,207–$311,158
- Hardware market price
- $403,907–$418,867
- Typical use
- Enterprise-grade production deployment (original China-compliant model, now discontinued)
H20 96GB
96 GB · 4,000 GB/s · Export-compliant special supply card (significantly reduced compute power)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 4,000 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 6,800 tok/sin 4,810 tok/s · many concurrent requests
- Hardware list price
- n/a
- Hardware market price
- $95,741–$131,644
- Typical use
- Suboptimal choice in export-compliant environments, medium-to-large scale inference deployment (FP16 compute power is only about 15% of H100, but higher memory bandwidth offers an advantage in decode-intensive scenarios)
A800 PCIe 80GB
80 GB · 1,935 GB/s · Compliant special supply card (export compliance status adjusted multiple times)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 1,935 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 1,974 tok/sin 6,084 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $119,676–$124,463
- Hardware market price
- $129,250–$143,612
- Typical use
- Medium-scale inference/training deployment in export-compliant environments
Ascend 910C 128GB
128 GB · 3,200 GB/s · Huawei dual-chip data center card (domestically produced)- Suggested precision
- INT8
- GPUs needed
- 8
- Est. decode speed
- 1,600 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 2,720 tok/sin 24,440 tok/s · many concurrent requests
- Hardware list price
- n/a
- Hardware market price
- $95,741–$239,353
- Typical use
- Domestically produced large model training and inference clusters, CloudMatrix and other super-node solutions
Ascend 910B 64GB
64 GB · 400 GB/s · Domestic alternative (not subject to US export controls)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 400 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 680 tok/sin 10,400 tok/s · many concurrent requests
- Hardware list price
- n/a
- Hardware market price
- $131,644–$143,612
- Typical use
- Domestic alternative, not subject to US export controls, medium-to-large scale inference/training deployment, memory bandwidth significantly lower than NVIDIA cards of similar VRAM capacity
RTX PRO 6000 Blackwell 96GB
96 GB · 1,792 GB/s · Professional workstation/server card (China-specific version is 6000D 84GB)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 1,792 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 1,828 tok/sin 9,750 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $128,200Cloud rental (live) ×8: $9.62–$18.15/hr (2026-11-05)
- Hardware market price
- $140,691
- Typical use
- Single-card large VRAM inference, local deployment of medium MoE models, private all-in-one machine
RTX PRO 5500 Blackwell 84GB
84 GB · 1,398 GB/s · Professional workstation card (specs are NVIDIA preliminary data)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 8
- Est. decode speed
- 1,398 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 1,426 tok/sin 8,814 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $91,147
- Hardware market price
- $101,725–$107,709
- Typical use
- Workstation large VRAM inference, departmental private deployment
RTX PRO 5000 Blackwell 48GB
48 GB · 1,344 GB/s · Professional workstation card- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 16 (exceeds one node: multi-node needed)
- Est. decode speed
- 2,688 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 2,742 tok/sin 10,062 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- n/aCloud rental (live) ×16: $8.98–$17.10/hr (2026-11-05)
- Hardware market price
- $129,250
- Typical use
- Small to medium model inference, development testing, single-machine multi-card private deployment
L40S 48GB
48 GB · 864 GB/s · Compliant inference card- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 16 (exceeds one node: multi-node needed)
- Est. decode speed
- 1,728 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 1,763 tok/sin 14,118 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $107,709–$155,579Cloud rental (live) ×16: $9.62–$14.59/hr (2026-11-05)
- Hardware market price
- $107,709–$155,579
- Typical use
- SME teams building inference services, medium-low concurrency
L20 48GB
48 GB · 864 GB/s · Compliant inference card- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 16 (exceeds one node: multi-node needed)
- Est. decode speed
- 1,728 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 1,763 tok/sin 4,660 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $55,882–$56,202
- Hardware market price
- $83,773
- Typical use
- SME teams building inference services, medium-low concurrency
RTX 5090 32GB
32 GB · 1,792 GB/s · Consumer flagship (not sold in China, 5090D for domestic market)- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 16 (exceeds one node: multi-node needed)
- Est. decode speed
- 3,584 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 3,656 tok/sin 16,341 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- n/aCloud rental (live) ×16: $7.07–$8.51/hr (2026-11-05)
- Hardware market price
- $107,706
- Typical use
- Local inference and development for individuals/small teams, multi-card VRAM pooling for quantized MoE
RTX 5090D 24GB
24 GB · 1,344 GB/s · Consumer-grade/small-scale self-built- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 32 (exceeds one node: multi-node needed)
- Est. decode speed
- 5,376 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 5,484 tok/sin 23,166 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $78,982Cloud rental (live) ×32: $14.14–$16.96/hr (2026-11-05)
- Hardware market price
- $134,037–$191,482
- Typical use
- Personal development/small-scale verification/POC, not recommended for production-level high concurrency
RTX 4090D 24GB
24 GB · 1,008 GB/s · Consumer-grade/small-scale self-built- Suggested precision
- INT4 (AWQ/GPTQ)
- GPUs needed
- 32 (exceeds one node: multi-node needed)
- Est. decode speed
- 4,032 tok/ssingle-stream theoretical ceiling
- Est. batched throughput
- out 4,113 tok/sin 22,932 tok/s · many concurrent requests · PCIe multi-GPU penalty
- Hardware list price
- $62,227–$90,949Cloud rental (live) ×32: $10.82–$13.95/hr (2026-11-05)
- Hardware market price
- $124,463–$143,612
- Typical use
- Personal development/small-scale verification/POC, not recommended for production-level high concurrency
FAQ
FAQ
How is the estimated decode speed calculated?
Estimated speed = (memory bandwidth × GPU count) ÷ (active parameters × bytes per parameter): the memory-bound upper limit for one request (roofline). It scales linearly with GPU count and ignores interconnect overhead, batching and framework losses, so real results are usually well below it.
Can I trust the speed numbers for consumer GPUs?
The linear assumption is reasonable for NVLink data-center cards (H/B series) but too optimistic for consumer PCIe cards without NVLink (RTX 5090D/4090D); multi-GPU consumer setups lose far more to communication, so equal-looking numbers do not mean equal performance.
How is “Suggested precision” chosen?
It is the highest precision that fits within 8 GPUs; beyond that it is flagged as multi-node.
Where do the prices come from?
USD converted at ¥6.68 per USD. Cloud rental is the international vast.ai market rate, not a China purchase price.