Zen Go: deepseek-v4.1-flash decodes ~30% slower than the same model on Command Code and DeepSeek's own API #51010

Activity

1. tigerwho1888 commented on Oct 15, 2026

   

   tigerwho1888

   on Oct 15, 2026

   Author

   More actions

   Three weeks on, this route has got worse rather than better, so here are fresh numbers using the same method (output tokens ÷ the message's own stream window, from my client's records):

   period	steps	median tok/s	p10
   09-17 – 09-18 (launch week)	4,258	254	183
   09-24 (the report above)	1,184	183	110
   10-13 – 10-14	5,223	119	70

   End to end — what the user actually waits for, TTFT included — a step now medians 80 tok/s against 124 at launch, and about a quarter of the last two days' steps come in under 60 tok/s.

   The peer routes on the same machine over 10-12 – 10-14 are unchanged: Command Code 177 median (n=2,972) and DeepSeek's own API 194 (n=27). So it still looks specific to the Go gateway rather than the model or my connection.

   A controlled probe today — same prompt, 600 output tokens, no context — gave 132 / 143 / 186 tok/s, and one run at 44 tok/s with a 16.2s TTFT, so the route also stalls now rather than just decoding slowly.

   Still hoping for an answer on whether Go-route requests are served on different capacity from the pay-as-you-go Zen endpoint. Happy to run the identical-prompt A/B against both endpoints any time if that helps.
2. neriousy commented on Oct 21, 2026

   

   neriousy

   on Oct 21, 2026

   Member

   More actions

   Thanks for the report. This is about how the model itself behaves when served through OpenCode Go (repeating itself, broken tool calls, speed, or rejecting large requests), not a bug in the OpenCode app. We pass these reports to our upstream providers. Starting a new session or running /compact usually helps with loops and oversized-request errors. Closing this to keep the tracker focused; the main discussion is in #50206.
3. neriousy

   closed this as not plannedon Oct 21, 2026
4. LOFWagner commented on Oct 26, 2026

   

   LOFWagner

   on Oct 26, 2026

   More actions

   Its now sitting at sub 40 tok/s for me, anyone else having the same performance?

   👍React with 👍3Reacted by EduardoVizioli, Amr Rady and Leonid
5. Wulgaren commented on Oct 28, 2026

   

   Wulgaren

   on Oct 28, 2026

   More actions

   Adding data to this, since the Go vs free capacity question here never got an answer.

   DeepSeek V4.1 Flash feels slow in OpenCode. The token speed is fine. The problem is the wait before each step produces anything, and it adds up over many agentic steps. I measured it from my local session DB instead of guessing.

   Setup: Poland (CET), macOS, OpenCode 2.2.0, 15–28 Oct 2026, 149 sessions, ~14k DeepSeek steps over three routes. TTFT here means the time from when a step is created to its first reasoning or tool part, so it includes OpenCode's own request setup. Error steps are excluded.

   route	TTFT median / p90 (s)	whole model stream median / p90 (s)	output tok/s (median)	steps with TTFT > 10 s
   DeepSeek API direct (deepseek/deepseek-flash, 1,405 steps)	1.5 / 2.4 (max 6.3)	3.5 / 12.7	~90	0.0%
   OpenCode Go (opencode-go/deepseek-v4.1-flash, 8,169 steps)	3.1 / 6.5–9.1 (max 184)	5.9–6.1 / 16.8–19.0	69–87	5.7%
   OpenCode free (opencode/deepseek-v4.1-flash-free, 4,500 steps)	2.4–2.5 / 6.2–7.8 (max 249)	5.0–5.9 / 13.7–29.7	72–98	6.1%

   • Decode speed is similar on every route. The difference is the first-token wait and its long tail.
   • At matched context size (~99% cache hits), the median TTFT is 1.7 s direct vs 3.0 s on Go at 150–300k tokens, and 2.2 s vs 4.0 s above 300k. Go's TTFT grows with context, the same pattern as opencode-go: deepseek-v4.1-flash first-token latency scales with context (p90 ~25s at 657k, minutes at the tail), plus empty completions and missing finish_reason #57055.
   • On Go, steps with a model stream over 30 s are 4% of steps but take 25% of all model time. The mean model time per step is 9.4 s on Go vs 5.9 s direct.
   • Tools aren't the bottleneck: the shell median is 0.1 s. The median prompt runs 9 steps, and ~92% of turn wall time is model time.
   • Go TTFT is barely better than the free tier, with the same ~6% of steps waiting more than 10 s.

   Questions:

   1. Regions: the Go docs say DeepSeek V4.1 Flash / V4 Pro need the Global region (the default allowed regions are Europe and US), and DeepSeek V4 Flash on suddenly requires "Enable models hosted in China" for OpenCode Go subscription #39845 shows "only available hosted in China". When will DeepSeek be served from the US or Europe? From Europe, that would likely cut TTFT a lot.
   2. Go vs free capacity: would you separate OpenCode Go (paid) traffic from free-tier traffic for DeepSeek, so Go users get faster, steadier responses?

   Related: #57055, #57849, #57403, #58003, #57270, #55579, #39845, #50155. I can share the aggregation query and anonymised timing rows if that helps.