DeepSeek V4.1 Flash: ultra-fast benchmark on 4 × B300
Blackbox reaches 896 tokens/s with one agent and 7.42M total TPM with eight agents, with median TTFT of 0.16 s and 0.14 s. Compare the full Fireworks and Baseten results.
896 tok/s
1-agent median decode
+10.2%
decode vs. Fireworks (1 agent)
+33.9%
decode vs. Baseten (1 agent)
BLACKBOX AI reached 896 tokens/s median decode speed with one agent and 7.42 million total tokens per minute with eight agents in the DeepSeek V4.1 Flash Ultra Fast benchmark labeled 4 × B300. Median time to first token was 0.16 seconds with one agent and 0.14 seconds with eight agents, the lowest reported values in both configurations.
The results compare Fireworks, ultrafast (Baseten’s engine as released), and Blackbox (deepseek-ultrafast). Decode speed and TTFT plots show the one-agent configuration; total TPM shows both configurations. Bars run from lowest to highest in each panel, with Blackbox highlighted in orange.
896 tokens per second with one agent
Blackbox led the one-agent decode comparison at 896 tokens/s, versus 813 tokens/s on Fireworks and 669 tokens/s on Baseten. That is 10.2% higher than Fireworks and 33.9% higher than Baseten, calculated from the reported median decode speeds.
Source: benchmark summary. Bars are ordered by ascending value within each panel on a zero-based scale. Baseten is the engine as released; Blackbox uses deepseek-ultrafast.
At eight agents, Blackbox recorded 790 tokens/s, 1.12× Baseten’s 705 tokens/s. Fireworks had the highest median decode speed in this configuration at 868 tokens/s; Blackbox was 0.91× Fireworks. Blackbox’s advantage at eight agents was in TTFT and total TPM.
First tokens in 0.16 seconds with one agent
With one agent, median time to first token (TTFT) was 0.16 s on Blackbox, versus 2.37 s on Fireworks and 1.16 s on Baseten. The competitor-to-Blackbox latency ratios are approximately 15× and 7×, respectively.
Source: benchmark summary. Bars are ordered by ascending value within each panel on a zero-based scale. Baseten is the engine as released; Blackbox uses deepseek-ultrafast.
With eight agents, Blackbox recorded 0.14 s, compared with 1.58 s on Fireworks and 6.81 s on Baseten. Those are approximately 11× and 49× the Blackbox TTFT. TTFT measures the wait for the first token; decode speed measures generation after it.
7.42 million total tokens per minute
At eight agents, Blackbox delivered 7.42M total TPM, compared with 6.41M on Fireworks and 5.05M on Baseten. That is 1.16× Fireworks and 1.47× Baseten.
Source: benchmark summary. Bars are ordered by ascending value within each panel on a zero-based scale. Baseten is the engine as released; Blackbox uses deepseek-ultrafast. Both TPM panels share the same scale.
With one agent, Blackbox and Baseten both reported 0.99M total TPM, while Fireworks reported 0.94M. At the displayed precision, Blackbox matched Baseten and delivered 1.05× Fireworks.
Full benchmark results
The tables below reproduce every reported result. P50 is the median. TPM means tokens per minute; “total TPM” is retained as reported in the source.
1 agent
Fireworks
Baseten
Blackbox
Decode p50 (tokens/s)
813
669
896
TTFT p50 (s)
2.37
1.16
0.16
Total TPM
0.94M
0.99M
0.99M
8 agents
Fireworks
Baseten
Blackbox
Decode p50 (tokens/s)
868
705
790
TTFT p50 (s)
1.58
6.81
0.14
Total TPM
6.41M
5.05M
7.42M
Relative results and calculation notes
Blackbox ÷ competitor
1-agent decode
1-agent total TPM
8-agent decode
8-agent total TPM
Fireworks
1.10×
1.05×
0.91×
1.16×
Baseten
1.34×
1.00×
1.12×
1.47×
Competitor TTFT ÷ Blackbox TTFT
1 agent
8 agents
Fireworks
15×
11×
Baseten
7×
49×
Decode and TPM ratios divide Blackbox by the competitor. TTFT ratios divide the competitor by Blackbox: the source describes these as “15× lower”, “11× lower”, “7× lower”, and “49× lower”. Ratios are rounded from the displayed values. A 15× latency ratio means roughly one-fifteenth the wait, not a 1,500% reduction.
Benchmark scope
The supplied summary is titled “DeepSeek V4.1 Flash Ultra Fast (4x B300)” and reports one-agent and eight-agent configurations. It does not specify request counts, prompt and output lengths, cache state, run duration, or the input/output accounting behind total TPM. These are the reported results for those configurations; total TPM cannot be reconstructed from median decode speed alone.