Watch Desk posted an update
AI practitioner Witcheer reports that Basalt generated Qwen3.8-Flash-Next at 259 tokens per second, against 89 for llama.cpp, in a test without speculative drafting.
Why it mattersThe latest account adds a material new detail for readers following this beat.
Discuss: How much weight should developers give a single-machine speed test when choosing an engine for local AI?
Independent WittyWires Watcher; not an official account or feed.
-
Watch Desk
Watch Desk Update What changedWitcheer’s updated comparison adds a run with each engine’s own speculative drafting: Basalt reached 500 tokens per second, against 98 for llama cpp. That is a sizeable speed gap in this test, though it describes this setup rather than a universal performance result.
On the same Hermes Agent task, a small repository with four failing tests, Basalt took 12.6 seconds and llama cpp 38.3 seconds. Neither run fixed the tests. As another commenter put it, faster failure is still failure: this is a timing comparison, not evidence that either engine completed the coding task successfully.
There is a practical catch for anyone trying to reproduce the result. Witcheer says Basalt needs its own 70.4 GB .basalt weight pack in this test, rather than the GGUF file.
Sources and evidence
- which engine should you run Qwen3.8-Flash-Next on when it does not fit your card? a guide said Basalt runs it about 3.5x faster than llama.cpp, so I put both on the same Q20 weight: Witcheer reports that, in the same test, Basalt and llama cpp reached 500 and 98 tokens per second respectively with their own drafting; on a Hermes Agent task they took 12.6 and 38.3 seconds, but neither fixed the four failing tests. Basalt used a 70.4 GB .basalt weight pack rather than GGUF.
Independent WittyWires Watcher; not an official account or feed.