Deepseek V4.1 flash TPS is very slow.

[XXXE88]

The TPS output from Deepseek V4.1 is very slow. Last week, it was averaging around 250, but today it's only around 100. This is almost a 50% decrease, which significantly impacts the user experience.

⚡ 107 tok/s · avg 128

[XXXE88]

And today's (September 28th) DeepSeek V4.1 Flash is significantly different from the official version.

[XXXE88]

GOAT Package; Intelligence is very low, with a severe decline; No feedback channels

[XXXE88]

The latency is also extremely high.

[centre009]

Reproducing the TPS/latency reports above with the transport metadata —
deepseek/deepseek-v4.1-flash is no longer being served by DeepSeek's own API.

Same key, same CLI channel (POST /alpha/generate), same model, tiny prompt.

Baseline (mid-September, archived):
provider-metadata → providerMetadata.deepseek {promptCacheHitTokens,
systemFingerprint} + providerMetadata.gateway.routing
{resolvedProvider:"deepseek", finalProvider:"deepseek",
credentialType:"system", only:["deepseek"]}

2026-09-29 00:49 +08:00:
provider-metadata → providerMetadata = {"novita": {...}}
no deepseek block, no gateway.routing at all.
Response headers changed too: x-fusion-transport: embedded,
x-ratelimit-*, and one response carried a Domain=novita.ai cookie.

Control in the same run: deepseek/deepseek-v4-pro still returns the full
DeepSeek-official metadata (finalProvider:"deepseek", systemFingerprint,
zero-markup cost reconciliation). So this is a flash-tier change, not
"Command Code left DeepSeek".

That lines up with the reports in this thread: ~250 → ~100 tok/s, much higher
latency, and output that is "significantly different from the official
version". Novita does list DeepSeek V4.1 Flash in its own model library.

Caveat: this is a single observation window right after the 2026-09-28
DeepSeek API outage, so it could be failover rather than a permanent switch.
Worth re-checking over the next days — the test is cheap and deterministic.

Why nobody noticed earlier: the official-vs-third-party routing only shows up
in the raw stream's provider-metadata event. Normal CLI/desktop UI surfaces
(tok/s, usage, cost) look identical either way.

[XXXE88]

hahaha