You wake from cryo 187 years late. ๐
StarStranded is a pixel-art action RPG I'm building solo in Godot.
1- Crash on a harsh world and survive.
2- Talk to anyone in your own words. The AI dialog engine runs on your PC.
3- Then choose: help the planet, or cleanse it.
Devlog starts today.
#indiedev
@CoffeeBeansDev Rejections are rough but you put yourself out there, that's the hard part. Go back to the why. You wanted a fun game with friends, and that's still worth making. Do what makes you feel good, life is too short.
@ChristianSelig Good to know it's not just the Duo. Is it only the layout around the island acting up, or are Live Activities in it affected too? About to find out with my app lol
@Recretos Cool to see AIFS lining up with GFS on this one. How often does AIFS split from the regular ECMWF this far out on strat stuff? And which do you lean on when they disagree?
Type anything to an NPC in StarStranded and they answer dynamically, in character, & out loud!
Everything runs on your own PC/SteamDeck. No cloud, no account.
1- llama.cpp running Phi-3.5-mini (Q4_K_M, about 2.4 GB)
2- Piper TTS for the voices, with pitch and distortion per character
3- whisper.cpp if you'd rather talk than type
4- About 29 tok/s on a SteamDeck.
Here's Unit 734, the drone waiting by your cryo pod, on why it was watching you sleep:
#screenshotsaturaday
You wake from cryo 187 years late. ๐
StarStranded is a pixel-art action RPG I'm building solo in Godot.
1- Crash on a harsh world and survive.
2- Talk to anyone in your own words. The AI dialog engine runs on your PC.
3- Then choose: help the planet, or cleanse it.
Devlog starts today.
#indiedev
Been experimenting with local models over the last weekโฆI had no idea they had come so far. Same RTX 5090, same 10 coding tasks, two Qwen 3.8 models. ๐ฅ๏ธ
27B (dense, NVFP4 on Infernix) vs Flash Next (125B MoE, about 6B active, IQ3_S on Strata).
1- Short prompts: the 27B is faster. 322 vs 245 tok/s.
2- At 120K+ context it flips. Flash Next holds 253, the 27B drops to 192.
3- Both passed 10 of 10. Flash Next finished in 14.2 min, the 27B in 42.8.
Iโll post those custom benches later.
The part I didn't expect was Strata.
A 125B model doesn't fit in 32 GB of VRAM. On one request Strata filled 31.1 of 31.5 GB, put another 32.8 GB in shared system memory, and still kept the GPU at 90%.
It's the first runtime where it feels like I'm using the whole machine, not just the card.
5090, Ryzen 9 9950X3D, 64 GB DDR5. @coldniko
@mitansh_j07 Appreciate it! NPCs feel so alive when you can speak to them in your voice and they reply with theirs. Itโs all local so the conversations never leave your PC/SteamDeck๐. More on how it works in the devlog soon.
Same struggle here. I ended up giving people both. Normie Mode is a fixed simple view, and if you want more you can turn cards on and off and drag them into your own order. Turning Normie off brings your layout back. Curious if you think that's a cop out vs nailing one default lol
@PatrickVaughn4 Appreciate you saying 'if the HRRR is correct.' Do you lean on it more for fog than for rain timing? Working on a weather app and i'd like to suggest various models that are known for being more accurate for certain variables.
@saurabhg A Live Activity that outlives the rain is exactly what I'm trying to avoid in the weather app I'm building. What would you have wanted/expected it to do instead? Happy to DM a TestFlight invite if you want to try mine.