Stanislav Trifan

[Stanislav Trifan @stanislav_3fun]

@HermesAgentTips Qwen3.8

[Stanislav Trifan @stanislav_3fun]

@atmoio Do local one. On high usage it just buys itself in about few months. And use the Claude for planing and critical checks

[Stanislav Trifan @stanislav_3fun]

@draginol It sounds sarcastic. Such a model does not scale well on two, so one is more than enough for good performance

[Stanislav Trifan @stanislav_3fun]

@draginol you compare 7k$ with half its price. A gpu with 24-32VRAM (AMD/NVIDIA - even some old ones like 3090) can do more than 100 tps(some even more than 200) of the 27b and the flash-next

[Stanislav Trifan @stanislav_3fun]

@0xSero Either I don't get it, but what is the actual throughput tested, to support the claim? Long context is really good. Would be helpful few numbers about the speed too

Thanks

[Stanislav Trifan @stanislav_3fun]

@meijer_s Buy the least capable hardware that can run the qwen3.8-27b wit more than 30tps decode and you would be surprised. Worth combining local with cloud LLMs, while not sure about anthropic. From time to time they look like scammers to me (codex, Claude, grok, https://t.co/QhKBtqCO6i)

[Stanislav Trifan @stanislav_3fun]

@lmstudio it stops to be accessible by the pi coder after update, I thought its intentionalbeing.+ remote access broke after last two updates (it seesm like it overrides the settings or something)

[Stanislav Trifan @stanislav_3fun]

Your updates broke the local setup, so thanks for motivating me to not rely on Bionic anymore. Thanks; there are a bunch of other options on the market.

Pushing so hard to cloud-only aggressively only makes your product less usable. Either way, there are plenty of choices.

The only downside is that it says a bunch about your mindset by breaking existing setups. And makes me move today rather than later. Was a short ride of a good product so far

[Stanislav Trifan @stanislav_3fun]

@arjunkmrm Looks like the TLA+ should grow in popularity along with the LLM tooling, too

[Stanislav Trifan @stanislav_3fun]

@designwkarthick Buy 64GB or more if it is possible (fully understandable that it might be expensive), but the 16gb of RAM for the last few years was a crime against humanity from Apple.

[Stanislav Trifan @stanislav_3fun]

Wonder if for most coding activities the qwen3.8-27b feels more predictable (and even faster on a consumer GPU) than having grok 4.7, or is it just for me?

[Stanislav Trifan @stanislav_3fun]

the 128gb ram mac is cheaper by ~2k, its not an exact substitution but is substantially faster 2x+ , and for 2k difference from this one you get the 256 version that is about 4 times faster. Mac is just faster GB/$, and for equal money I would buy a linux machine anyday and that was true until now

[Stanislav Trifan @stanislav_3fun]

Shall we stop calling everything AI? JEV is a classifier (a good one), but a classifier. LLM is LLM - and not an AI either; more generic, overall good, yet not AI, etc

[Stanislav Trifan @stanislav_3fun]

@droidbuilds So you need 28 years of running it to get to 0

[Stanislav Trifan @stanislav_3fun]

@Ananth7e Are you saying me, that it's worth 10 years of max subscription of any provider the hardware alone (to not count electricity and maintenance), and telling us that is cheap? High likely in just few years you will want better hardware etc

[Stanislav Trifan @stanislav_3fun]

@testerlabor It doesn't mean to be better just because of bigger numbers. Although I all for a good competition.

[Stanislav Trifan @stanislav_3fun]

@mr_r0b0t @NVIDIAAI Cool, does it have MTP support?

[Stanislav Trifan @stanislav_3fun]

A follow-up from Pergunta Insights: what would make “Why Skipping the Plan Is a False Economy” easier to act on? https://t.co/d7CbvEqxeP