Karol Spark

AI
Local LLMs
DGX Spark

[Karol Spark @_karol_spark]

Guess Who’s Back?

[OpenClaw🦞 @openclaw]

OpenClaw 2.0 has arrived
https://t.co/52iWXJQvCy

[Karol Spark @_karol_spark]

I'm getting many questions about a small plugin that I've built in the Hermes desktop app, which shows the actual generation tok/s. Not a benchmark, not a false figure – just realistic work with real tasks.

You can easily spot how this number compares to the promises from the recipe.

Happy to see this on your screen?
@Teknium

[Karol Spark @_karol_spark]

So the limits are dropping by almost 17%.
When you run a local AI model, your limit stays the same – no limits.

[ClaudeDevs @ClaudeDevs]

Starting September 14, we're permanently raising standard weekly limits in Claude Code by 25% for Pro, Max, Team, and seat-based Enterprise plans. Until then, the current 50% increase will be in place.

[Karol Spark @_karol_spark]

@l1inear @MiaAI_lab @Teknium maybe an idea?

[Karol Spark @_karol_spark]

@OwlCoreAI @MiaAI_lab no mate, this looks amazing! Should work for you really fine :) you will see when is slowing down and when is speeding. Better thing than benchmarks

[Karol Spark @_karol_spark]

@MiaAI_lab @Zai_org I can confirm. My test with $20 was failure as well

[Karol Spark @_karol_spark]

@OwlCoreAI @MiaAI_lab it's really helpful. You can see realistic generation. Which is sometimes far away from benchamarks :) Good luck!

[Karol Spark @_karol_spark]

@MiaAI_lab I build this with GLM-5.3 Flash - just asked to build plugin and add this to Hermes bottom right window.
Is showing in real time tokens generation and total token generated for current session

[Karol Spark @_karol_spark]

@0xSero @Tech2Wild 100% right, and I think best for 4 sparks now is EXL3 3bit or EXL/TR3 3.42 - same quality as NVFP4 with room for context

[Karol Spark @_karol_spark]

@u1tra_instinct Legend! Waiting for more details!

[Karol Spark @_karol_spark]

And who said that open-weights models are far behind?
New Terminal-Bench 4.0
Opus 5 🥇
Fable 5 🥈
GLM-5.3 🥉

Well done @Zai_org!

[Zixuan Li @ZixuanLi_]

Terminal-Bench 4.0 just dropped.

Benchmark iteration is catching up with model development.

[Karol Spark @_karol_spark]

@Tech2Wild Morning Chef

[Karol Spark @_karol_spark]

@t0nil0 @matteocollina @NVIDIAAI Didn't try this yet. Maybe at some point will try to replace them :) are you planning to do this?

[Karol Spark @_karol_spark]

@t0nil0 @matteocollina @NVIDIAAI tbh all depends from model and recipe = CPU and GPU usage. When I'm using some small models like GLM-5.3 flash or DeepSeekV4 flash fans are stable - literally no noise. Full GLM-5.2 or GLM-5.3 (dint' try this yet) noise can be pretty high while processing prompts.

[Karol Spark @_karol_spark]

@t0nil0 @matteocollina @NVIDIAAI CSR812 and you?

[Karol Spark @_karol_spark]

Sounds right?

[Karol Spark @_karol_spark]

@TechMDAI EXL3 is really solid! 98-99% quality and faster on sparks than NVFP4

[Karol Spark @_karol_spark]

@bridgemindai You are blinded… you are talking about here and now. Wait till December this year and tou posts will change dramatically.

While people like you are complaining… other ones already well prepared with equipment for running locally big models.

Trust me