钽TANDIU钽铥
SELF-HOSTING

K2 Horizon 375B A23B: Self-Hosting Estimate

K2 Horizon 375B A23B (IFM(MBZUAI)) on different GPUs: configuration, estimated performance and use cases.

Data updated: 2026-09-16

← All self-hosting estimates

Total parameters379BSets weight VRAM
Active parameters23BSets decode compute and speed
Public benchmarksNo public benchmark yet
Vendor API output speed—measured by Artificial Analysis; compare with local estimates below

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
4priced as 1 × 8-GPU HGX node
Est. decode speed
696 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 1,183 tok/sin 25,435 tok/s · many concurrent requests
Hardware list price
$400,084–$440,093Cloud rental (live) ×4: $39.76–$50.00/hr (2026-11-20)
Hardware market price
$2,545,253–$2,994,415
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
FP16/BF16 (full precision)
GPUs needed
8
Est. decode speed
1,391 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 2,365 tok/sin 56,522 tok/s · many concurrent requests
Hardware list price
$239,745–$359,617
Hardware market price
$718,660–$1,437,319
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
FP16/BF16 (full precision)
GPUs needed
8priced as 1 × 8-GPU HGX node
Est. decode speed
1,391 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 2,365 tok/sin 50,870 tok/s · many concurrent requests
Hardware list price
$360,075–$440,093Cloud rental (live) ×8: $66.22–$70.01/hr (2026-11-20)
Hardware market price
$449,162–$524,023
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
FP16/BF16 (full precision)
GPUs needed
8priced as 1 × 8-GPU HGX node
Est. decode speed
835 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 1,419 tok/sin 22,360 tok/s · many concurrent requests
Hardware list price
$248,053–$320,067Cloud rental (live) ×8: $29.90–$57.06/hr (2026-11-20)
Hardware market price
$1,003,129–$1,033,073
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
INT8
GPUs needed
8priced as 1 × 8-GPU HGX node
Est. decode speed
1,165 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 1,981 tok/sin 22,360 tok/s · many concurrent requests
Hardware list price
$296,061–$309,064Cloud rental (live) ×8: $13.77–$34.86/hr (2026-11-20)
Hardware market price
$501,565
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
INT8
GPUs needed
8priced as 1 × 8-GPU HGX node
Est. decode speed
1,165 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 1,981 tok/sin 22,360 tok/s · many concurrent requests
Hardware list price
$293,453–$311,419
Hardware market price
$404,246–$419,218
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
INT8
GPUs needed
8
Est. decode speed
1,391 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 2,365 tok/sin 3,346 tok/s · many concurrent requests
Hardware list price
n/a
Hardware market price
$209,609
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
INT8
GPUs needed
8
Est. decode speed
673 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 687 tok/sin 4,232 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
$119,777–$124,568
Hardware market price
$158,105
Typical use
Medium-scale inference/training deployment in export-compliant environments

Ascend 950DT 144GB

144 GB · 4,000 GB/s · Huawei's new-generation SuperNode card (domestic, parameters and prices for reference only)
Suggested precision
FP16/BF16 (full precision)
GPUs needed
8
Est. decode speed
696 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 1,183 tok/sin 11,304 tok/s · many concurrent requests
Hardware list price
n/a
Hardware market price
$929,467
Typical use
Domestic large model inference SuperNode (64-card SuperPoD), report states inference performance is approx. 0.86x that of H200 at the same latency

Ascend 910C 128GB

128 GB · 3,200 GB/s · Huawei dual-chip data center card (domestically produced)
Suggested precision
FP16/BF16 (full precision)
GPUs needed
8
Est. decode speed
557 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 946 tok/sin 17,002 tok/s · many concurrent requests
Hardware list price
n/a
Hardware market price
$95,821–$239,553
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
INT8
GPUs needed
8
Est. decode speed
139 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 237 tok/sin 7,235 tok/s · many concurrent requests
Hardware list price
n/a
Hardware market price
$131,754–$143,732
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
INT8
GPUs needed
8
Est. decode speed
623 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 636 tok/sin 6,783 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
$128,307Cloud rental (live) ×8: $10.58–$16.06/hr (2026-11-20)
Hardware market price
$203,620
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
INT8
GPUs needed
8
Est. decode speed
486 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 496 tok/sin 6,131 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
$91,223
Hardware market price
$101,810–$107,799
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
8
Est. decode speed
935 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 954 tok/sin 3,500 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
n/aCloud rental (live) ×8: $4.82–$9.09/hr (2026-11-20)
Hardware market price
$34,467
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
8
Est. decode speed
601 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 613 tok/sin 4,911 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
$53,899–$77,855Cloud rental (live) ×8: $3.74–$6.42/hr (2026-11-20)
Hardware market price
$53,899–$77,855
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
8
Est. decode speed
601 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 613 tok/sin 1,621 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
$27,964–$28,125
Hardware market price
$41,922
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
8
Est. decode speed
1,247 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 1,272 tok/sin 5,684 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
$16,002Cloud rental (live) ×8: $3.65–$4.81/hr (2026-11-20)
Hardware market price
$57,493–$62,284
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
16 (exceeds one node: multi-node needed)
Est. decode speed
1,870 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 1,907 tok/sin 8,058 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
$39,524
Hardware market price
$67,075–$95,821
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
16 (exceeds one node: multi-node needed)
Est. decode speed
1,402 tok/ssingle-stream theoretical ceiling
Est. batched throughput
out 1,430 tok/sin 7,976 tok/s · many concurrent requests · PCIe multi-GPU penalty
Hardware list price
$31,140
Hardware market price
$41,919
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.