Latent Local

@RealZairaonX Does this only work with a sub? Or can you use this with a pay-as-you-go API? What about if it's out of credits? lol

V4-Pro feels practical for real dev tasks. Close to Opus on SWE-bench, ~7x cheaper, and supports 1M context. Still trails on harder reasoning, but strong ROI. Free to try here: zenmux.ai

@ZenMuxAI @deepseek_ai Links are broken for me...

@deepseek_ai 🔗DeepSeek: DeepSeek V4 Pro (Free)：zenmux.ai/deepseek/deeps… 🔗DeepSeek: DeepSeek V4 Flash (Free)：zenmux.ai/deepseek/deeps…

DeepSeek-V4 @deepseek_ai is HERE FREE on @ZenMuxAI ! 🚀 API & Chat are free for a limited time – jump right in! Two editions, one mission: 🔷 V4-Pro – max performance, built for agentic coding. Real‑world feel: better than Sonnet 4.5, close to Opus 4.6 (non‑thinking). ⚡ V4-Flash – smaller, faster, lower cost. Simple agent tasks? On par with Pro. 🔥 One API, all protocols – OpenAI Chat/Responses + Anthropic Messages. Use it inside Codex, Claude Code, OpenCode, etc. PAYGO & Builder Plan already supported. Try for free now → @ZenMuxAI #DeepSeek #DeepSeekV4 #AI #LLM #VibeCoding #OpenSource #ZenMuxAI

@mil000 I can get about 1hr-ish with adversarial agent reviews and synthesis. No clue how anyone is having agents run for HOURS and not just getting pure slop… The context window alone. What am I missing?

Can anyone explain to me what in gods name they need long running agents for? I haven’t seen a single reason/ use case so far

@dexhorthy I dunno. I want to read a 6mo article about “X JUST DROPPED….AND IT DESTROYS” Or that local models are just as good as SOTA models. Better even but 1Mx cheaper.

hey Nikita I really wanna see more ai written hype slop about harness engineering can you help me out

@AniC_dev @asciidotdev How hands off is it? What do you use?

imagine having your CTO agent manage 10 parallel cloud claude code/codex instances to work on your repos in autonomy and produce hosted fullstack dev builds for each new feature you just asked him while you euromaxx with your based polish gf my life rn

Hey, @sama codex usage dried up. I'm stuck using Pi (thinking off) for the last 2%. Reset limits! Unless you reply, then I'll upgrade to $200 plan, lol.

@jarekceborski Probably depends on setup. Speed I think matters but then you’re not talking a MacBook, you’re talking $$$ GPUs…

What’s a good use case for local LLMs, except privacy concerns?

@typecraft_dev Anyone not saying XPS is lying. That hidden trackpad look is slick.

Dell xps 14 or framework 13 pro. Looks wise. What do you like better?

@scaling01 Everything is still so early. We should all recognize that when there’s so many players in the market and it’s moving so fast. Telltale sign it’s early. When everyone starts settling around “the real path” for X (agents/harness/hardware etc) then we’ll know things are slowing.

November 2025 changed everything for me Claude 4 and GPT-5 weren't really amazing and my expectation was that we are likely in a slower timeline. But then Opus 4.5 came out. Since then every new model and revenue report has just reinforced this new world-view that we are still early and everyone is constantly underestimating it Predictions that were ridiculously bullish a year ago, 6 months ago or even 3 months ago were underestimates If you want to call the top on the AI bubble, then go ahead, but to me this seems like shorting Moore's Law in 1970 As long as revenue and compute keeps growing it seems delusional to try to call a top

I spent 1.7B tokens on Codex (Pro 5x) and 80M tokens on Claude Code (Max 20x) yesterday. Guess which one gave me a usage-limit warning.

@LLMpsycho @Metzjosh2 I don’t think A CLI proxy would work…

@Metzjosh2 @latentlocal i use cli proxy but is not stable,byok errors many times,is there any alternative?

droid computers is the kind of thing that makes you wonder why agents didn't have persistent homes from day one. keeps your setup across sessions, no more rebuilding deps from zero. byom is free

@fejau_inc *AND* they’re ALL growing haha

Unlike the dot com bubble there’s actually revenues This will upset many people

@sheriyuo You can’t say things like that if the gap is GROWING…. If it was shrinking your comment makes more sense. Not saying DS4 team isn’t incredibly competent, but they’re not SOTA level (maybe for OS only…)

Nothing wrong with that. 8 months is not a meaningful gap at this pace of progress. Distillation is a powerful tool and it benefits every AI giant building models today. Do not underestimate the DeepSeek team, they are genuinely formidable and moving fast.

@kimmonismus But you’re not playing the other side of the equation. Look at their models: They’re falling further and further behind US AI. A not insignificant piece of that is chip superiority. So, what would you have us do?

I think many people still don't grasp what this means. Chris Miller analyzed in his book "Chip War" that China is about 10 years behind in lithography machine production. Meanwhile, Huawei chips are achieving roughly 80% of the performance of the NVIDIA H100. China is probably more like 2 years behind the US, and the gap is narrowing. I will be writing a detailed analysis on this soon because it's far more significant than the models themselves.

@luciascarlet If you’re wondering how much better the $100 is than the $20 on codex…it’s legit. Genuinely feels like you’ve got breathing room. If you’re running codex solo, it’ll last. If you’re plugging all day in parallel…ehh it’ll get sucked up still. Still hit my weekly!

is ChatGPT Pro $100 any good I probably won’t get it but maybe if I end up using Codex more and burn my Plus limits… hmph…

@hieuSSR They didn’t *just* start, c’mon. And no, have you signed up for an API code yet?

Cloudflare just started hosting AI models, would you switch?

@gmi_cloud @mosh_Ontong @ollama Do you post your model speeds?

@mosh_Ontong @ollama just use us

DeepSeek V4 Pro vs Xiaomi MIMO V2.5 Pro MIMO V2.5-Pro wins by: 1. ~1 min faster than DeepSeek-V4-Pro 2. cleaner aesthetic Our reasoning: MIMO V2.5 Pro is natively multimodal, so it has a stronger sense of what good visuals look like. Curious to see how DeepSeek V4 Vision will change that

@enesakar I feel ya. It’s like someone dumped a bunch of cool parts on the floor and they’re like: “y’know, you could build some pretty neat stuff with this!” I’m sure I could! But man…seems like a lotta work first haha

openclaw is great. but too complex for me. too many configs, hard to update etc. i will write a simpler version of it for my own needs as my weekend project.

@LLMpsycho Can you use subs in Droid?

Setup update : Droid with Codex $100 + Opencode Go