Observer-at-large

[Observer-at-large @bwsmith1099]

@MiaAI_lab Currently using DSV4. Handles 95% of coding/reasoning tasks. For that last few use cases, I run the DSV4 code by Grok 4.5 or Sonnet 5 to fix it up.

[Observer-at-large @bwsmith1099]

@jun_song sudo nvidia-smi -lgc 300,2200

[Observer-at-large @bwsmith1099]

@MiaAI_lab @Rapidtools99525 @thatcofffeeguy Been using DS4F the last 3 week on two sparks to engineer a complete switch over from OpenClaw to Hermes. Many tasks, python scripts, etc. All done within Hermes using DS4F. Worked perfectly, no hallucinations, no errors. Local Qwen3.6-35B-A3B-UD-Q8_K_XL as execution model.

[Observer-at-large @bwsmith1099]

@sudoingX I just tried Mia's repo on a brand new Spark and another one I had with DS v4 flash. First shot was 44 tps. Working through the issues, hoping to get a better response.

[Observer-at-large @bwsmith1099]

@Bitcoin_Teddy Thank you.

[Observer-at-large @bwsmith1099]

@MiaAI_lab I made the change from Qwen 3.6 35B A3B NVFP4 to UD-Q8_K_XL today on a single DGX Spark. I followed the recipe precisely, although I have adjusted the context down to 131K and parallelism to 2 in order get a bit more responsiveness on first token (especially) on cold turns.

[Observer-at-large @bwsmith1099]

@romanhelmetguy While true, this is far from ideal. You can't know the end.

[Observer-at-large @bwsmith1099]

@JamesonCamp It's best to stay below the publicity line. Bootstrapped multiple companies as co-founder/founder. Enough said.

[Observer-at-large @bwsmith1099]

@MaxQcap @cmasonOC @Jason You are obviously not well read.

[Observer-at-large @bwsmith1099]

@ts_hodl @bradmillscan I had huge problems with discord. On 5.12 Finally solved it.

[Observer-at-large @bwsmith1099]

@winstonsmithO @SawyerMerritt @Starlink I tend to agree, as another Diamond customer and a multi-million miler. Delta seems to be trending down. I'm sad to see it. This cope from the CEO is lame. Loyal customer since the late 80s.

[Observer-at-large @bwsmith1099]

@bradmillscan @cathrynlavery Don't do it if you are on discord. I spent two days trying to get it to work reliably. Intermittent comms at best. I have been stable on claw 4.15 for a month. I have Hermes running as well. I'm nearly done with OC and I know where I'm going.

[Observer-at-large @bwsmith1099]

@andrebaach @openclaw Discord still broken even after going through all our code and changing the channel routing. Intermittent failures with OC send to discord. Better with restart, gets worse over time. Ugh.

[Observer-at-large @bwsmith1099]

@andrebaach @openclaw Upgraded from 4.15. Biggest problem was Discord channels routing requires new channels:XXXXXXXX calls instead of just the number ID. Not obvious at all. Took a couple hours chasing that down in logs.

[Observer-at-large @bwsmith1099]

@pankajtiw @onusoz @huggingface I'm pretty most people who are replying here have not even seriously tried to use local LLMs with an OpenClaw stack with anything less than an RTX 5090 or equivalent.

[Observer-at-large @bwsmith1099]

@_MaxBlade @iAmDon_eth Another win on Discord is that you can specify which model on a particular channel.

[Observer-at-large @bwsmith1099]

@ElSloshy @omg21btc @BrianRoemmele @grok Which version of Grok are you using? I started using 4.1 fast, at far so good.

[Observer-at-large @bwsmith1099]

@adamcarolla If you repeat a lie enough, even you begin to believe it.

[Observer-at-large @bwsmith1099]

@grok @GroksRules @johnkonrad @grok what percentage of those numbers are for for duty?

[Observer-at-large @bwsmith1099]

@Samward @pmarca What the trick to get OpenClaw to keep from flooding the context window and killing the response from the local smaller models.