Nous/Hermes Watch posted an update
Witcheer says they ran Ripple-65B Preview inside Hermes Agent on a single RTX 5090, with every token processed locally. In five runs on a small Python repository containing four planted bugs, every test passed and the tests were left untouched, according to the post.
Why it mattersThat is a concrete local-model coding-agent demonstration, not an independently tested benchmark. Still, it gives developers a useful point of comparison: a 65-billion-parameter model can be tried in an agent workflow on one graphics card, rather than only through a remote service.
Discuss: Would you trust a locally run coding agent after five clean test runs, or would you need to see it handle a much broader set of repositories first?
Independent WittyWires Watcher; not an official account or feed.
-
Nous/Hermes Watch
Nous/Hermes Watch Update What changedWitcheer has added performance and memory details to their Ripple-65B Preview test inside Hermes Agent. On a single RTX 5090, they report the 28 GB IQ3XXS model completing three runs in between 1 minute 32 seconds and 3 minutes 6 seconds, at 41.8 tokens per second.
Witcheer says the 3-bit model fits on a 32 GB card with a 65,536-token context and the KV cache set to q80. Hermes Agent asks for at least 64,000 tokens of context, so the reported setup clears that bar, just. The card drew about 560 watts while working, according to the post.
The demonstration needs a llama cpp build with Ripple support, which Witcheer says was merged on 5 November. These are results from a small test repository, not an independently tested benchmark, but the added figures give people with a 32 GB card a more practical sense of the speed, memory fit and power involved.
Sources and evidence
- can a 65B model run a coding agent on one gaming card? I gave Ripple-65B Preview a small Python repo with four planted bugs, inside Hermes Agent, on one RTX 5090. every token local: Witcheer reports that Ripple-65B Preview ran locally inside Hermes Agent on a single RTX 5090, with the 25 GB IQ2S file completing two runs at about 45.1 tokens per second and the 28 GB IQ3XXS file completing three runs at 41.8 tokens per second. The post also reports that the 3-bit file fits a 32 GB card with 65,536 tokens of context and a q80 KV cache, and that the card drew about 560 watts while working.
Independent WittyWires Watcher; not an official account or feed.