Hacker News: XCSme

New comment by XCSme in "Qwen 3.8 27B is excellent, but it defaults to overthinking things"

XCSme — Mon, 17 Aug 2026 07:14:14 +0000

Also, note that the Luna was run and costs were calculated through OpenRouter, via API. With a ChatGPT subscription it's likely even cheaper.

Also, Qwen 3.7 27B is actually Terra level.

New comment by XCSme in "Qwen 3.8 27B is excellent, but it defaults to overthinking things"

XCSme — Mon, 17 Aug 2026 07:12:37 +0000

I removed the extra links to sources for the electricity prices, but that's the average cost in EU, where I live.

https://ec.europa.eu/eurostat/web/products-eurostat-news/w/d...

New comment by XCSme in "Qwen 3.8 27B is excellent, but it defaults to overthinking things"

XCSme — Mon, 17 Aug 2026 06:58:09 +0000

My comparison of its reasoning efforts[0] seems to show that it only really supports 3 modes: none, low, xhigh.

Low and medium are basically the same.

Also, the electricity it costs to run on a 3090 is not negligible, so that it's cheaper to use Luna high via API than Qwen 3.8 27b locally, hardware costs excluding.

[0]: https://aibenchy.com/compare/qwen-qwen3-8-27b-high/qwen-qwen...

New comment by XCSme in "Qwen 3.8 27B"

XCSme — Sat, 15 Aug 2026 12:30:15 +0000

It's an amazing model, it's GLM-5.2 level[0], running locally...

I tested it on my 3090, took like 8 hours to benchmark it and my room became a furnace (35+ deg outside temp), but it's really good.

Now, in theory, you can talk directly to your computer and tell it what to do, and it does everything locally.

[0]: https://aibenchy.com/compare/qwen-qwen3-8-27b-medium/z-ai-gl...

New comment by XCSme in "Qwen 3.8 27B"

XCSme — Fri, 14 Aug 2026 23:38:20 +0000

Msi afterburner, you go around 900mv curve editor, raise it up to normal clock frequency, and it uses like 260w instead 300w for same performance.

There's some YouTube guides for it.

I also undervolted my new 5070ti, same tdp, around 260w instead of 300w and like 8% better performance.

New comment by XCSme in "Qwen 3.8 27B"

XCSme — Fri, 14 Aug 2026 19:31:26 +0000

5900x, 3090 24gb (slightly undervolted), 128gb ddr4, running via Ollama.

I am benchmarking it now locally, will put the results and speed/tps on aibenchy.com

New comment by XCSme in "Qwen 3.8 27B"

XCSme — Fri, 14 Aug 2026 19:10:09 +0000

Oh, ok, that's like the average tps for most AI providers

New comment by XCSme in "Qwen 3.8 27B"

XCSme — Fri, 14 Aug 2026 18:59:57 +0000

Why slow? I see ~50tps on a single 3090

New comment by XCSme in "Qwen 3.8 27B"

XCSme — Fri, 14 Aug 2026 18:59:01 +0000

With default config via Ollama and 65k context I get 50tps on a 3090.

New comment by XCSme in "Gemini 3.7 Flash"

XCSme — Thu, 13 Aug 2026 18:56:09 +0000

They both have horizontal scroll on mobile...

New comment by XCSme in "DeepSeek V4 Pro 0813"

XCSme — Thu, 13 Aug 2026 10:49:25 +0000

So where does the guarantee stop? At the GPU driver level? Firmware level?

What if the GPU has a custom bios flash that somehow logs the unencrypted prompts?

New comment by XCSme in "Pixel Watch 5"

XCSme — Thu, 13 Aug 2026 10:23:56 +0000

Why do I feel like there haven't been any hardware releases since the Macbook Neo? Market is so slow, probably due to the DRAM shortages.

No new GPUs, CPUs, technologies, etc...

New comment by XCSme in "Grok 4.6"

XCSme — Thu, 13 Aug 2026 10:12:30 +0000

Yeah, Sol models have higher in/out cost but are incredibly token efficient.

Also, in those tests Sol Low did better, but you can also compare the price vs Sol High, then it's getting a bit closer.

So Grok 4.6 is still not the best choice when paying API rates, but they are improving fast.

Also, the more important difference is that sol is a lot faster.

New comment by XCSme in "DeepSeek V4 Pro 0813"

XCSme — Thu, 13 Aug 2026 03:30:40 +0000

Open source doesn't matter if someone else is running it, right? They can change it?

As long as the prompt is not encrypted at some point, and I don't think LLMs can run on encrypted prompts, then it can be read.

New comment by XCSme in "Nvidia Nemotron 3.5 Lightning and NeMo Switchyard"

XCSme — Thu, 13 Aug 2026 01:19:37 +0000

This is not a solution for now, just a moonshot solution for maybe 5-10 years from now.

There's no immediate fix for the RAM supply issue, but as long as it's manufacturing issue, not a raw materials issue, time always solves it.

New comment by XCSme in "Grok 4.6"

XCSme — Thu, 13 Aug 2026 00:12:57 +0000

Grok 4.6 vs Sol 5.6 vs Opus 5:

https://aibenchy.com/compare/openai-gpt-5-6-sol-low/x-ai-gro...

New comment by XCSme in "Grok 4.6"

XCSme — Thu, 13 Aug 2026 00:10:20 +0000

Does really well and ~2x cheaper than Qwen3.8 2.4T, they have same pricing but grok is around 2x more token efficient:

https://aibenchy.com/compare/qwen-qwen3-8-2-4t-a95b-low/x-ai...

New comment by XCSme in "Qwen3.8-2.4T"

XCSme — Thu, 13 Aug 2026 00:07:12 +0000

Interestingly, the high variant does a lot worse and failed to generate a valid SVG (and the low variant use more tokens than the high one, so maybe their reasoning efforts are not working properly).

The solar system animation is also the coolest looking I've seen, unfortunately the animation doesn't work:

https://aibenchy.com/compare/qwen-qwen3-8-2-4t-a95b-low/qwen...

New comment by XCSme in "Qwen3.8-2.4T"

XCSme — Thu, 13 Aug 2026 00:02:31 +0000

That's a really cool hamster [0], unfortunately it's really expensive now, 2x more expensive than Grok 4.6[1].

[0]: https://aibenchy.com/compare/x-ai-grok-4-6-high/bytedance-se...

[1]: https://aibenchy.com/compare/x-ai-grok-4-6-high/bytedance-se...

New comment by XCSme in "DeepSeek V4 Pro 0813"

XCSme — Wed, 12 Aug 2026 23:38:12 +0000

I don't understand how this works?

Is it another proxy on top? What stops the provider from reading/storing the prompts at the LLM execution level?