TotallyArea

@SquaredCubeRBX sick, let’s connect on discord (totallyarea). i’m also looking to start releasing some models as well (messing w a qwen 3.5 9B to release publicly and learn more about training).

@TotallyArea this was released like 4 months ago so it was rented back then (h100/h200s on vast.ai) newer versions are being trained on my own compute now :)

hey, roblox devs! Here's a set of specialized coding model I developed @ Cybus for roblox game dev: you can run it on around 2-3 rtx 3090s, or a single 6000 pro, or 2 4090/5090s. (also an older version with quantizations available that fit on smaller gpus) links in the comments:

this is no replacement for good programmers or frontier ai, but it’s good for free & private access to something solely on your computer. the cost is electricity (minimal!!) use for light/medium or repetitive tasks, frontier/programmer for anything else

@berko_k i’ve been using it at around 64k+ context on gpu alone, can push to 128k with cpu offloading. you can also put your cache on a quant as well which really helps!

@TotallyArea hell yeah, would be cool. although I can't even get myself to switch to Luna for daily coding tasks. I don't know if I can bare using a 16-32K context window

DEVS, would you be interested in a SPECIALIZED AI model to assist with coding? not a wrapper, API or subscription. download and run it for 100% FREE, 100% PRIVATE. small enough to run on even on a laptop, designed to help with everyday lua tasks. 👇👇 Oppoah Type 9B (Preview)

good question, huggingface.co/spaces/Torpedo… i’ve been using this leaderboard to try and base my stuff off of, but i dont know the exact specs/settings used on everything. one thing that kind of makes this leaderboard a bit scuffed is some tests may have been ran offline vs api (api can search the web). both of my tests were taken offline, im sure would do even better with search. every ai model will sometimes give a wrong or outdated answer. i’ve been learning a lot about how models actually output their answer. the goal of the tiny model is just to use it on repetitive tasks, for people who might not be able to get a subscription, or would like to experiment (like me). in regards to a system prompt to steer claude, i would either spend some time writing what its job is, and to maybe verify what methods it uses, or ask chatgpt for a prompt (it’s frankly good at such).

@TotallyArea How does Fable 5/Opus 4.8? perform on that same test? To be honest I would be more interested in an open source skill for Claude that has all up to date APIs, syntax and tricks on Studio. Claude sometimes uses outdated stuff or switches to internet to look up modern syntax, etc.

#RobloxDev #roblox

would love to hear feedback from devs and if this is of interest. questions, comments, concerns, let me know.

some stats ongoing: 3000 question multiple choice test was applied to the fine tuned model, and finished with an 85% accuracy, and a failure-to-answer (ai death loop) of just 0.7%. ~$1.12 in electricity, 9 hours. 3.6m output tokens base model is estimated to take approximately 18 hours, with an accuracy of 71-74%. estimated to output at least 9 million tokens. base model testing is underway, and is at a 75% completion for its first test, hence where we estimate our numbers for above.

the goal is to build something that everyone can use and benefit from. it won’t replace claude or codex on complicated tasks, but it can save you meaningful usage and time on easier or repetitive tasks. (you can use it on claude’s terminal with ease, allowing it directly into studio!)

over the past couple of weeks, i’ve been working on filtering through tons of high quality, private data and publicly licensed data. we removed anything that was outdated, deprecated or just generally low quality.

@frrazers would be interesting to have one that could run locally and offline👀

Vibe coded my own thumbnail generator in 10 minutes because I'm not paying £20 a month for a Gemini wrapper

@dunik_7 i’ve done this with glm 5.2 (colibri) and it’s just horrifically slow (<0.6tps) and the model was just straight hallucinating. cool but early tech, not very useful in its current state.

he made 405-billion-parameter models run on a single 8GB gaming GPU. instead of turning it into a paid service, he open-sourced the whole thing. Gavin Li Founder and CEO of Anima AI in the Bay Area. 'airllm' - it keeps one layer on the GPU at a time and streams the rest, so a 70B model fits in 4GB and DeepSeek's full 671B fits in about 12GB 22,700 stars. Apache 2.0. → github.com/lyogavin / he has 808 followers. / his repo has 22,700 stars. bookmark it. that's what shipping something impossible looks like.

@HuggingModels i ran this on my computer with the intel ultra core 9 285k and 64gb ddr5 ram, it ran approx 0.3-0.7TPS. the model hallucinates really bad for some reason though (on many responses).

Big news for AI on a budget. GLM 5.2 Colibri int4 is a Mixture of Experts model that runs entirely on your CPU. No GPU needed. It's fast, efficient, and opens up advanced language capabilities to anyone with a standard computer. This is a game changer for offline AI.

@Nikola_Devv ptr and d7 will be the downfall

Too early to give up? From peaks of almost 2k ccus, now cant be stable at 1k ccus.. overnight #Roblox #RobloxDevs

@RhysiebroTV @l_eiif data loss supposedly.

what happened to Slime RNG? @l_eiif

@SurferRBLX roblox has been cracking down (and wrongly) on every game recently for some reason. ive had countless emails over age rating rejected on xyz.

One of my games that maintained a consistent CCU was recently taken down for “personal information.” We’ve already contacted support to try and get it reinstated, but I’m worried this could completely kill the game’s playerbase. It was generating steady income that my brother and I rely on to pay developers and advertise our other games. What makes it even more frustrating is that the game had been up for over 2 years with 15M+ visits and never had issues until now. I really hope Roblox considers some form of compensation for the downtime, or at least provides a way to help recover the playerbase once the game is restored. Super frustrating situation. #Roblox #RobloxDevs #RobloxDev

@intrance @EvanZir i have 30 ccu games making more than games with 10k ccu, verified with analytics.

@EvanZir well this isn't even true but ok

Most people on Roblox don’t realize this… There are hundreds of 300 CCU games out there quietly making more money than a game with 5,000 or 10,000 CCU. Speaking from experience 😅 Monetization > CCU It’s just about creating a game for the right niche

@NVIDIAGeForce GeForce Season Need to play bf6 with a bf6 themed 5090

❄️ #GeForceSeason of Giving ❄️ Want your chance at a custom wrapped Battlefield 6 GeForce RTX 5090?! Comment “GeForce Season” below and tell us why you deserve this Holiday upgrade 🎄