Massed Compute

Those are two different renter decisions. Blackwell is not the LFM throughput card. The MiniCPM-V-4.7 BF16 file is 65.59 GiB, so it does not fit a 48 GB card.

Live Massed Compute list rates on 19 October 2026 are the rates in the tables. L40S is $0.97/hr. A100 SXM4 is $1.38/hr. RTX PRO 6000 Blackwell is $2.19/hr. H100 is $2.73/hr. H100 NVL is $3.11/hr. The L40S list rate rose from $0.88 to $0.97 on 8 September 2026. These pages use $0.97. Confirm current GPU pricing before you launch.

The numbers that decide the card

Two jobs. Value and speed are not the same GPU, and on MiniCPM-V-4.7 only one GPU finished.

LFM2.5-24B-A2B (LiquidAI/LFM2.5-24B-A2B, vLLM 0.32.0, random 128/128 at concurrency 32). The BF16 checkpoint is 44.65 GiB. L40S serves it with online FP8. A100 and Blackwell serve BF16. L40S delivers 1110.4 tok/s per dollar. A100 delivers 1415.3 tok/s, the throughput high in this set. Blackwell delivers the lowest median first token, 101.9 ms, and the worst tokens per dollar. The three rows do not share quantization or KV dtype, so the gaps are indicative.

MiniCPM-V-4.7-35B-A3B (openbmb/MiniCPM-V-4.7-35B-A3B, BF16, 65.59 GiB). The timed requests are text completions. The server left video enabled and profiled two maximum-size video items. Blackwell finished at 1018.8 tok/s and $0.597 per million output tokens. A100 and H100 loaded 65.65 GiB, then exited before the bench with a negative KV-cache budget.

LFM2.5-24B-A2B: packed tokens

Same checkpoint and revision. Same vLLM 0.32.0. Random 128 in / 128 out, ignore-eos, concurrency 32, prefix cache off, --max-model-len 8192. Capture date 19 October 2026. L40S is online FP8 with fp8 KV. A100 is BF16 with bf16 KV. Blackwell is BF16 with fp8 KV. Treat the gaps as indicative, not an A/B of one kernel. Method and showcases: LFM2.5-24B-A2B and MiniCPM-V-4.7 GPU bench (PR).

Least expensive launched card is gpu_1x_l40s. Throughput high in this set is gpu_1x_A100_SXM4. First-token card is gpu_1x_pro_6000_blackwell. Cheaper 48 GB cards were in stock on 19 October 2026 and were not launched: gpu_1x_a6000 $0.57/hr, gpu_1x_a6000_low_ram $0.55/hr, gpu_1x_6000_ada $0.79/hr, gpu_1x_l40 $0.86/hr. gpu_1x_a100 at $1.35/hr had no capacity. Those SKUs are untested. L40S is the least expensive card we ran, not a cheapest-fit claim.

L40S is the value card in this set. A100 is the packed-throughput card. Blackwell is the first-token card. Those three jobs are not the same GPU.

What extra money buys on A100

A100 is the c32 throughput high: 1415.3 tok/s versus 1077.1 on L40S. The hourly is $1.38 versus $0.97. Tokens per dollar fall from 1110.4 to 1025.6. Cost per million output tokens rises from $0.250 to $0.271. Median first token moves from 182.7 ms to 178.8 ms. That is 3.9 ms. Pay A100 when this queue is tok/s-bound and 1077.1 tok/s is the bottleneck. The first-token gap is not the purchase.

What extra money buys on Blackwell

Blackwell is 1378.7 tok/s. That is under the A100 row. The hourly is $2.19. Tokens per dollar are 629.5, the low in this set. Cost per million output tokens is $0.441.

The purchase is first token. 101.9 ms versus 182.7 ms on L40S and 178.8 ms on A100. Around 100 ms after send reads as instant. Pay Blackwell when someone is waiting on the first word and 182.7 ms is the problem. Do not pay it for tokens per dollar, and do not pay it to beat A100 on packed tok/s. In this set it does not.

The BF16 weights are 44.65 GiB. L40S shows 44.99 GiB total, so the published L40S row is online FP8. nvidia-smi while the server was up held 41.12 GiB of 44.99 GiB on L40S, 73.36 GiB of 80.00 GiB on A100, and 87.60 GiB of 95.59 GiB on Blackwell (MiB / 1024). The first L40S start left the attention backend unset, picked FlashInfer, and died because the image has no nvcc. The published row forced Triton and ran at 8192.

Concurrency 1 output tok/s: L40S 115.7, A100 136.3, Blackwell 152.2. Each c32 row completed 160 prompts and failed 0.

MiniCPM-V-4.7: one finished card

Same vLLM 0.32.0 profile: random 128/128, concurrency 32, --max-model-len 8192, prefix cache off. BF16 weights, 65.59 GiB. The timed requests are text completions. The server used --limit-mm-per-prompt '{"image": 1}', which caps images and leaves video enabled. The logs show an encoder-cache budget of 16384 tokens, profiled with two maximum-size video items. Capture date 19 October 2026. Method: same PR.

gpu_1x_pro_6000_blackwell is the only launched card that held this model on that profile. Its c32 output throughput is 1018.8 tok/s, 465.2 tok/s per dollar, $0.597 per million output tokens. Concurrency 1 was 148.0 tok/s. The c32 row completed 160 prompts and failed 0. nvidia-smi while the server was up held 82.78 GiB of 95.59 GiB.

Both launched 80 GB cards failed the same way before the bench. Each log records model loading took 65.65 GiB. Allocator OOM warnings start once encoder profiling begins, before the KV budget is computed. Available KV memory is then −3.38 GiB on gpu_1x_A100_SXM4 at $1.38/hr (bf16 KV) and −2.51 GiB on gpu_1x_h100 at $2.73/hr (FlashAttention, fp8 KV). Both exit with ValueError: No available memory for the cache blocks. Those OOM warnings are not from allocating KV. With a negative KV budget, a shorter --max-model-len is not the retry these logs point at. The 16384 batched-token/encoder budget and --gpu-memory-utilization 0.92 were not retried.

The $1.38/hr A100 was not tested with --limit-mm-per-prompt '{"image":1,"video":0}'. gpu_1x_h100_nvl (94 GB) at $3.11/hr was in stock and was not launched. gpu_1x_h200_nvl had no capacity. 48 GB cards were not launched. The BF16 checkpoint is 65.59 GiB, so this file does not fit a 48 GB card. These captures leave the image-only profile and the 94 GB H100 NVL untested. They do not establish that 96 GB is required.

This is a text-completion bench on a vision-language model. Do not read 1018.8 tok/s as image or video generation speed.

Where speed and dollars disagree

How to launch

1. Sign in to the Massed Compute marketplace and open Deploy.
2. For LFM2.5-24B-A2B, pick gpu_1x_l40s. Take gpu_1x_A100_SXM4 when packed tok/s is the product. Take gpu_1x_pro_6000_blackwell when first token is the product.
3. For this MiniCPM-V-4.7 profile, the only finished launch is gpu_1x_pro_6000_blackwell.
4. Choose an Ubuntu image with NVIDIA drivers.
5. Add your SSH key. Launch.

Pay only for the hours you use. Confirm the live rate on GPU pricing at launch. The SKUs in the tables are the ones we ran on 19 October 2026.

Frequently Asked Questions