DeepSeek’s routed agent prototype finished a preliminary test in a 12.4-minute median. A single session with a sub-agent tool took 20.4 minutes. But when that session was told to delegate first, its median fell to 13.8 minutes.
The prototype also used more tokens. And the test used slow mock tool calls, so these times do not tell us how the agents perform when workers do real coding.
DeepSeek’s Router Finished the Test Faster
The prototype turned chat messages into work records, checked them against team policies, and sent accepted requests to workers with their own sessions and directories. Follow-up messages went straight to the worker already handling a ticket. DeepSeek found that removing this shortcut increased router decisions by about 30%.
The test involved about 300 chat and CI messages and roughly 20 work requests. DeepSeek reported a 15.7-minute time at the 95th percentile for the prototype, compared with 29.5 minutes for the single session. The lead was visible beyond the median, but it remains a preliminary task-time result.
That distinction matters. The tool calls were slow mocks rather than real coding work. The reported times show how the setups handled this test; they do not yet show whether the same lead holds when workers must write, run, and fix code.
A Delegation Prompt Narrowed the Gap
With a neutral prompt, the single session never called a sub-agent in five runs. Told to delegate first, it reached a 13.8-minute median. Its 95th-percentile time stayed at 29 minutes, close to the earlier 29.5-minute figure.
So the prompt changed the median comparison considerably, while the slower end of the results changed little. The routed prototype still had the lower reported median, at 12.4 minutes, but the gap depends in part on whether the single session actually delegates.
Token use adds another part to the comparison. The prototype used about 23 million cached tokens, against 6 million for the neutral session, though another analysis put that baseline higher. The reported cash estimate for the extra tokens was 5 to 10 cents at DeepSeek’s Flash rates on November 1st. The test did not identify which model it used, which limits how much that price estimate can explain about this particular run.
The released code also has clear operating limits. Self-identified work starts in observe mode, which logs decisions without starting tasks. Workers can run commands as the operating system user, including reading that user’s home directory.
DeepSeek describes a cloud connection and a watchdog for stuck workers, but neither appears in the released code. It also lists credential leakage into model context as unsolved. Those details matter if the prototype is being read as something ready to run, rather than as a preliminary timing result.
Strata’s Fast Figure Depends on the Model
The local inference numbers need the same care with their conditions. In Strata’s RTX 5070 figures, its details dock reports 93 tokens per second for Quen 3.8 Flash. Its lead table reports four to six tokens per second for the much larger Deepseek V4.1 Flash.
Those are figures for different models. The fast result for Quen 3.8 Flash does not describe the speed of Deepseek V4.1 Flash, even though both appear in Strata’s reported results.
Strata keeps frequently used model experts on the GPU and the rest in system RAM. When the model selects an expert that is not on the card, the CPU handles it. Strata’s documentation says the larger model has about 125 billion parameters, with about six billion active for each token, and needs about 34 GB of RAM for its experts plus overhead.
Memory placement made a large difference in an independent repeated-run comparison. Putting every expert in RAM instead of reading them from the SSD gave the biggest speed jump. In that comparison, an RTX 4080 reached about 49 tokens per second, while a DGX Spark running VLM managed about 38.
Prompt length and task change the picture again. Strata’s details figures were about a fifth slower at the longest context shown, and an issue collects hard-crash reports for prompts somewhere past 50,000 tokens. In one image-pointing comparison, Strata’s error was more than three times larger than llama.cpp’s.
A local speed figure needs its model, prompt length, and task beside it. DeepSeek’s timing lead needs its prompt and token use beside it too. Its next test is whether that preliminary lead holds when the workers do real coding.