SupraLabs/Supra-50M-Base
Text Generation • 51.8M • Updated • 1.74k • 29
Max Current
maxcurrent
AI & ML interests
None yet
Recent Activity
liked a model less than a minute ago
SupraLabs/Supra-50M-Base liked a model 2 minutes ago
SupraLabs/Supra-50M-Reasoning liked a model 3 minutes ago
SupraLabs/Supra-1.5-50M-Base-expOrganizations
None yet
liked a
model less than a minute ago
liked a
model 2 minutes ago
liked a
model 3 minutes ago
liked a
model about 20 hours ago
liked 5
models 1 day ago
empero-ai/Qwable-9B-Claude-Fable-5
Text Generation • 9B • Updated • 52 • • 9
tvall43/Qwen3.6-14B-A3B-FableVibes-GGUF
Text Generation • 14B • Updated • 4.09k • 14
WeiboAI/VibeThinker-3B
Text Generation • 3B • Updated • 1.95k • 314
squ11z1/Gravity-2
Text Generation • 3B • Updated • 9 • 1
squ11z1/Mythos-nano
Text Generation • 3B • Updated • 144 • 8
reacted to YMRohit's post with 🚀 1 day ago
Post
2977
A 1B model that writes GPU kernels you can trust
I fine-tuned OpenBMB's MiniCPM5-1B to write Triton GPU kernels, then let an immutable referee decide if they are real: compile, check correctness against PyTorch on adversarial inputs, time against eager, torch.compile, and torch.compile max-autotune, then block the known ways of gaming the benchmark.
The 1B setup beat torch.compile max-autotune in 12/12 independently seeded runs. The larger Qwen3.6-27B smith pushed the same referee loop further: 76 verified compiler-beating kernels on H200, with 69 surviving a 5-run stability gate and 7 kept as single-shot probes on unseen problems. On a 376-cell shape/dtype grid, the stability-gated kernels keep a 1.49x geomean, with about 10% of cells losing and reported per cell.
Honest bound: these are scheduling wins on memory-bound ops, not new algorithms or wins over cuBLAS/FlashAttention. The scarce thing is not the big model, it is the verifier it cannot fool.
Full write-up: https://huggingface.co/blog/YMRohit/ouroboros-kernel-mint
Try it: build-small-hackathon/ouroboros-kernel-mint
2-min demo: https://youtu.be/ViicZHktb-A
Built for #BuildSmallHackathon with MiniCPM, Qwen, Triton, Gradio, Codex, and Modal H200s.
I fine-tuned OpenBMB's MiniCPM5-1B to write Triton GPU kernels, then let an immutable referee decide if they are real: compile, check correctness against PyTorch on adversarial inputs, time against eager, torch.compile, and torch.compile max-autotune, then block the known ways of gaming the benchmark.
The 1B setup beat torch.compile max-autotune in 12/12 independently seeded runs. The larger Qwen3.6-27B smith pushed the same referee loop further: 76 verified compiler-beating kernels on H200, with 69 surviving a 5-run stability gate and 7 kept as single-shot probes on unseen problems. On a 376-cell shape/dtype grid, the stability-gated kernels keep a 1.49x geomean, with about 10% of cells losing and reported per cell.
Honest bound: these are scheduling wins on memory-bound ops, not new algorithms or wins over cuBLAS/FlashAttention. The scarce thing is not the big model, it is the verifier it cannot fool.
Full write-up: https://huggingface.co/blog/YMRohit/ouroboros-kernel-mint
Try it: build-small-hackathon/ouroboros-kernel-mint
2-min demo: https://youtu.be/ViicZHktb-A
Built for #BuildSmallHackathon with MiniCPM, Qwen, Triton, Gradio, Codex, and Modal H200s.
- 1 reply
liked 2
models 9 days ago
liked a
model 11 days ago
reacted to AxionLab-official's post with 🔥 13 days ago
Post
6584
We're happy to announce that we released a Reasoning tuned version of Supra-50M!
SupraLabs/Supra-50M-Reasoning
SupraLabs/Supra-50M-Reasoning
liked 3
models 13 days ago
liked a
model 14 days ago
liked a
model 16 days ago
liked a
model 17 days ago
