@yuhasbeentaken Current report from OpenCode itself is p50 2.6s TTFT and 143 TPS, which is way slower than DeepSeek, running at sub 1s TTFT and 275 TPS. I Don't know where you get your numbers from.
@thdxr During peak hours TTFT can be extremely high, more than a minute.
Also I noticed that since you cache per session, sending 2 identical prompts in 2 different sessions makes you miss it. May be worth checking.
@thdxr explore: 0%
Deepseek 4.1 Flash
Explore disabled bc it tends to mess up due to not reporting fine details. Better without.
The TTFT and TPS are deteriorating day by day. Sometimes waiting 40 secs for the first token, how is that even acceptable. I would pay more for lower TTFT.
@fanjiewang@mr_peanux TPS went from 230 down to 140 in 1 week or two.
TTFT has been all over the place.
I would like to know if it would be possible to be put in contact with someone for relevant feedbacks of your products. Would be really appreciated. Thank you.
@manongfu One thing most people forget is that an LLM has to be trained on those long conversation, that would make >1M harder to train as well. To be honest I think it's the sweet spot for now. Best quality/price to run.