filipe
@filicroval
data eng | 1xAsus Ascent GX10 | benchmarking local models so you don't have to
business inquiries: filipe@filicroval.com
266 Following 153.5K Followers
muse spark 1.3 is really good with 3D and three.js
it recreated Lies of P's mechanical heart from multiple screenshot references
it generated an exploded view too. so cool!
hy4 preview is definitely not a model we should sleep on
same prompt for each model, one-shot.
i would have expected the gap between both to be bigger but Hy4 Preview still comes close (it's twice the size but still). very huge leap since Hy3
curious which one you'd pick before checking the labels
Show more
i gave Ox Alpha the main menu from ECHO and asked it to recreate it
only ONE (1) prompt was enough to build this, using the reference
this is one of the most beautiful output i've ever had from any model
the fact that it's from a stealth model is mindblowing
Show more
no it's not, Ox Alpha is not an american model
the measured performance using the new OpenAI chip.
better performance per Watt, less latency and better interactivity.
they expect to deploy this chip in ChatGPT's infrastructure by the end of the year!
Show more
holy, new OpenAI chip is looking very strong in comparison to NVIDIA's chips
NVIDIA is already cooking Rubin Ultra next year and Feynman the year after, i wonder if Jalapeño will still be able to catch up
Show more
the measured performance using the new OpenAI chip.
better performance per Watt, less latency and better interactivity.
they expect to deploy this chip in ChatGPT's infrastructure by the end of the year!
Show more
holy, new OpenAI chip is looking very strong in comparison to NVIDIA's chips
NVIDIA is already cooking Rubin Ultra next year and Feynman the year after, i wonder if Jalapeño will still be able to catch up
Show more
plot twist: Qwen 3.8 Max just mogged them BOTH
same lighthouse prompt, and it delivered a way better landscape + lighthouse.
Qwen iterated for one hour until delivering this masterpiece
Sol was not even close, and even Ox Alpha's version looks basic next to this
don't sleep on this model
Show more
Ox Alpha built this very satisfying Powerwash Simulator
no idea what changed but the outputs were not THAT good yesterday
all assets generated from scratch: entire courtyard, the wall, the spray mechanics, the grime stripping off brick by brick
could it be a new checkpoint? i'm not really sure.
Show more
my agent harness tierlist 🤖
Claude Code, Codex, Hermes Agent and Oh My Pi in S+ i think most people would agree with that
i've used most of these at least once for testing.
what did I overrate or underrate?
make yours here:
Show more
holy, new OpenAI chip is looking very strong in comparison to NVIDIA's chips
NVIDIA is already cooking Rubin Ultra next year and Feynman the year after, i wonder if Jalapeño will still be able to catch up
Show more
my agent harness tierlist 🤖
Claude Code, Codex, Hermes Agent and Oh My Pi in S+ i think most people would agree with that
i've used most of these at least once for testing.
what did I overrate or underrate?
make yours here:
Show more
Ox Alpha built this very satisfying Powerwash Simulator
no idea what changed but the outputs were not THAT good yesterday
all assets generated from scratch: entire courtyard, the wall, the spray mechanics, the grime stripping off brick by brick
could it be a new checkpoint? i'm not really sure.
Show more
Sonnet 4.6 and Terra above Kimi K3, GLM-5.3 and DS4 flash is genuinely insane
GLM-5.3, Qwen3.8 Max and DS4 in C tier
that has to be ragebait
Best Model List
S+ tier - Fable, Sol and Opus 4.8
A - Sonnet 5.6, Terra
B - fun to experiment
C - cost-optimized, when u are on budget
D - whatever
Sonnet 4.6 and Terra above Kimi K3, GLM-5.3 and DS4 flash is genuinely insane
GLM-5.3, Qwen3.8 Max and DS4 in C tier
that has to be ragebait
Best Model List
S+ tier - Fable, Sol and Opus 4.8
A - Sonnet 5.6, Terra
B - fun to experiment
C - cost-optimized, when u are on budget
D - whatever
plot twist: Qwen 3.8 Max just mogged them BOTH
same lighthouse prompt, and it delivered a way better landscape + lighthouse.
Qwen iterated for one hour until delivering this masterpiece
Sol was not even close, and even Ox Alpha's version looks basic next to this
don't sleep on this model
Show more
plot twist: Qwen 3.8 Max just mogged them BOTH
same lighthouse prompt, and it delivered a way better landscape + lighthouse.
Qwen iterated for one hour until delivering this masterpiece
Sol was not even close, and even Ox Alpha's version looks basic next to this
don't sleep on this model
Show more
DGX Spark owners, rejoice!
excited to test it
SuperQwen3.8-27b-abliterated is now live and open for everyone 🚀
- Uncensored for freedom
- Broken weight from abliteration fixed with agents swarm
- fixed overthinking problem
- 1M context, Multimodal
Sorry about the late release.
Super-tune takes much more time than normal abliterated models, because fixing weights and eval takes about a week.
BF16/MLX/NVFP4 ⬇️
Show more
wake up, another @tonbistudio masterclass
Today's video is a full beginner's guide to Berd, the funky agent workspace for solo builders released by @blocks last week!
I did a quick setup demo last week, but this one digs in deeper to see everything you can do with Berd, and also tests its token efficiency compared to ClaudeCode. Check it out!
Show more
Xiaomi released a DGX Spark equivalent, that's huge
120B + 3B dual models, fast/slow switching, 1.22 TB/s near-memory bandwidth, nearly 4x the Spark. All on Xiaomi's own silicon at 150W.
More local AI boxes = more demand for open weights = more open weights get released. This flywheel is the best thing happening in AI right now. Your data stays home, your models never get deprecated, and no one can revoke your access.
Still a prototype, but the direction is very promising. Frontier models at home is slowly becoming reality
Show more
小米发了一个 AI 的本地的主机，搭载他们新发布的三个芯片，O3、O100 和 D100，支持 120B 和 3B 双模型，支持快慢系统的切换。
Xiaomi released a DGX Spark equivalent, that's huge
120B + 3B dual models, fast/slow switching, 1.22 TB/s near-memory bandwidth, nearly 4x the Spark. All on Xiaomi's own silicon at 150W.
More local AI boxes = more demand for open weights = more open weights get released. This flywheel is the best thing happening in AI right now. Your data stays home, your models never get deprecated, and no one can revoke your access.
Still a prototype, but the direction is very promising. Frontier models at home is slowly becoming reality
Show more
小米发了一个 AI 的本地的主机，搭载他们新发布的三个芯片，O3、O100 和 D100，支持 120B 和 3B 双模型，支持快慢系统的切换。
