# X account claims high t/s setup, but thin on details — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/c8b1b19541c593ca

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-09-28 12:26 ingested 2026-09-29 02:43

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1wsd7q9/x_account_claims_high_ts_setup_but_thin_on_details/)

## retained text



```
According to this post it is possible to reach very high numbers using mtp, however I have failed to reproduce the 50+ tps for 5060ti. 

Am I just ignorant or how exactly do this? Or is this a fake post?  
Freshly built llama fork for sm120 (Blackwell):   
[https://github.com/Anbeeld/beellama.cpp](https://github.com/Anbeeld/beellama.cpp)  
"C:\\llama\\build\\bin\\llama-server.exe" \^

  \-m "%MODEL%" \^

  \--port 8090 --host [127.0.0.1](http://127.0.0.1) \^

  \-ngl 99 \^

  \--cache-type-k kvarn3 --cache-type-v kvarn3 \^

  \--flash-attn on \^

  \--load-mode mlock \^

  \--jinja \^

  \-c 98304 --parallel 1 \^

  \--fit off \^

  \--spec-type draft-mtp --spec-draft-n-max 4 --spec-draft-ubatch-size 128 \^

  \-ctkd q8\_0 -ctvd q4\_0 \^

  \--kv-tail-tokens auto

Runs at 20-35 t/s. 
```



document fetched 2026-09-29 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1wsd7q9/x_account_claims_high_ts_setup_but_thin_on_details
