Learning Local AI | DGX Spark | Hermes Agent Local models + self-improving agents local.ai/localinference
Bangkok, Thailand
Based in Thailand
- Tweets 648
- Following 1,301
- Followers 353
- Likes 759
Apple must be getting crunched. I ordered an M5 Max MacBook Pro 128GB 8 weeks ago with a 6-8 week delivery estimate. They just called and said expect it the middle of October.
5
351
This has pretty much always been the case. Lot of love for OC since it put me on this path but reliability has always been a problem.
1
2
111
The pricing for DGX Sparks here in Thailand: $5,797, $5,689 and $7,676. No idea why the Asus box is so much more expensive. The MSA has a 3 year warranty (vs 1 year for the others) and has a Gen5 drive (vs Gen4 in the others).
1
9
1,872
If anyone has kids and they are interested in education in the AI era, this is an excellent resource and tool for both parents and kids. I have been looking through it and this deserves more visibility.
We have made some real progress on our SMF WisdomForge site this weekend. We continue to build out content, refine what we have up and make this a resource for all ages with AI integration baked in using age appropriate templated @NousResearch Hermes instances. Still have quite a way to go bit feeling much better about where we are with it. smfwisdomforge.com/
2
1
9
893
I agree, I have 4 agents in non-stop training runs for more than 36 hours straight using DSV4F and it is rock solid.
1
4
788
Benchmarked DeepSeek V4 Flash on 2× DGX Spark (TP=2): 36.63 tok/s on our StoryBench-compatible run and 34.35 tok/s on exact DharmaBench—about 22–27% faster than the GLM-5.3 Flash EXL3 result. Clean runs, no timeouts or CoT leakage.
Given that GLM has eyes and overall feels like a more capable model, I am thinking the trade off is worth it.
1
4
346
I really don't view it as a high percentage play to piss of the richest human on the planet and also, potentially, the smartest.
2
55
Hermes configs have now been changed to set reasoning to:
-Medium for normal Hermes agent work.
-preserve_thinking: false so old internal reasoning doesn’t accumulate across a long tool session.
-low for quick classification, routing, or routine tool calls.
-enable_thinking: false for trivial work where latency matters.
-xhigh only when deliberately requested for genuinely difficult reasoning.
11
1,868
New finding (to me at least)..Qwen3.8 in Hermes defaults to xhigh reasoning unless Hermes spefically specifies a different reasoning level. In Hermes at least, xhigh will make it think indefinitely without generating an answer...well tested up to 4.5 hours and 4M tokens burned purely thinking.
4
588
An update in using Hermes for everyday business tasks:
We are in the 3rd round of tuning a Hermes bot to specialize in building financial models. Currently, there is an orchestrator bot overseeing an analyst bot, both running DSV4F. After analyzing the output with Codex, we are moving the analyst bot to Qwen 3.8 27B to compare error catching efficiency.
I am guessing we will also wind up splitting the workload between 1 Hermes bot that is tuned to create assumptions pages, and another Hermes bot that is tuned to produce the model.
8
14
1,916