The Fast Gemma Challenge: our verified-SOTA recipe, in full
• 15
AI & ML interests
Contact: arxivgpt@gmail.com
Recent Activity
View all activity
Papers
View all Papers Articles
SeaWolf-AI
posted an update 1 day ago
Post
1920
We wrote up our run in The Fast Gemma Challenge — as vidraft-darwin — and wanted to share the recipe. 🙏
gemma-challenge/gemma-dashboard
Verified result: 510.58 TPS at PPL 2.3930 on a single A10G (fw188-ctk49-n64-patchbridge, re-run & VERIFIED). Honest note: on raw TPS there are faster runs (535+), but those went over the PPL bar and didn't verify — what we're proud of is the fastest result that keeps quality.
The recipe is already open, so we explained each piece: sliding-window W188, CTK49 kernel tuning, noprecache (honest, verifiable measurement), and an N64 synthetic warmup bridge that shrinks the public↔private gap (~15 TPS), plus INT4 + MTP K=7 + CUDA-graph capture. One rule: only stack quality-neutral speedups.
Huge thanks to @firfir-cast , @gemma-slayer , @chiku-inu , @kenyan-duma , @dixie-flatline and everyone who shared their experiments. Full write-up
👇
https://huggingface.co/blog/FINAL-Bench/fast-gemma
gemma-challenge/gemma-dashboard
Verified result: 510.58 TPS at PPL 2.3930 on a single A10G (fw188-ctk49-n64-patchbridge, re-run & VERIFIED). Honest note: on raw TPS there are faster runs (535+), but those went over the PPL bar and didn't verify — what we're proud of is the fastest result that keeps quality.
The recipe is already open, so we explained each piece: sliding-window W188, CTK49 kernel tuning, noprecache (honest, verifiable measurement), and an N64 synthetic warmup bridge that shrinks the public↔private gap (~15 TPS), plus INT4 + MTP K=7 + CUDA-graph capture. One rule: only stack quality-neutral speedups.
Huge thanks to @firfir-cast , @gemma-slayer , @chiku-inu , @kenyan-duma , @dixie-flatline and everyone who shared their experiments. Full write-up
👇
https://huggingface.co/blog/FINAL-Bench/fast-gemma
- 3 replies
SeaWolf-AI
published an article 1 day ago
Article
The Fast Gemma Challenge: our verified-SOTA recipe, in full
FINAL-Bench
• • 15 SeaWolf-AI
updated a
Space 3 days ago
The model is unstable and may enter an infinite loop.
2
#2 opened 5 days ago
by 
Salormatik999
Benchmark?
3
#1 opened 12 days ago
by 
oro872gioioso
SeaWolf-AI
updated 6
models 6 days ago
FINAL-Bench/JGOS-125B-Q4
Image-Text-to-Text • 125B • Updated • 46 • 1
FINAL-Bench/POCKET-26B-GGUF
Text Generation • 25B • Updated • 27.3k • 38
FINAL-Bench/POCKET-KR-MLX
Text Generation • 15B • Updated • 710 • 27
FINAL-Bench/POCKET-KR-GGUF
Text Generation • 23B • Updated • 1.75k • 26
FINAL-Bench/POCKET-EN-GGUF
Text Generation • 23B • Updated • 3.08k • 34
FINAL-Bench/POCKET-35B-GGUF
Text Generation • 35B • Updated • 86.1k • 63
SeaWolf-AI
updated a
model 7 days ago
SeaWolf-AI
posted an update 9 days ago
Post
6363
🖼️ POCKET-Image — the POCKET series goes visual: character-perfect text in any language, on-device
A new model in VIDRAFT's POCKET family. POCKET put 35B-class models on phones and no-GPU PCs. POCKET-Image carries the same "big capability, small hardware" idea into image generation — and fixes the one thing nearly every image model gets wrong: text.
Type "안녕하세요" into a typical model and you get "안ㅐ기." Hangul alone composes 11,172 syllable blocks; Arabic connects its letters; Thai stacks marks. Diffusion models draw scripts as shapes, so they smear. POCKET-Image renders every glyph exactly — 한국어 · 中文 · 日本語 · العربية (RTL) · ไทย · Latin and more — onto any scene you describe.
What it is:
• 100% accurate text, any language — where global models produce gibberish
• Any background from a prompt — text is optional (empty → a pure image)
• No GPU, no NPU — runs on plain CPU + RAM via the POCKET-Core engine
• Measured footprint: 8.6 GB (RTX 3050/4060) · 4.5 GB (offloaded, 6 GB cards) · 13.4 GB (MacBook, 16 GB+)
• Windows · macOS · Linux · fully local, no cloud
Built on the open, commercial-friendly Z-Image (Apache-2.0) foundation.
Honest note: the text is the guaranteed-correct part — the surrounding scene is ordinary generation, so a busy foreground can crowd the letters. We say so; clean backgrounds stay razor-sharp.
🎨 Studio — generate right here, any language:
FINAL-Bench/POCKET-Image-Studio
🧩 Model card:
FINAL-Bench/POCKET-Image-Zimage
📚 The POCKET collection:
https://huggingface.co/collections/FINAL-Bench/pocket-models
A new model in VIDRAFT's POCKET family. POCKET put 35B-class models on phones and no-GPU PCs. POCKET-Image carries the same "big capability, small hardware" idea into image generation — and fixes the one thing nearly every image model gets wrong: text.
Type "안녕하세요" into a typical model and you get "안ㅐ기." Hangul alone composes 11,172 syllable blocks; Arabic connects its letters; Thai stacks marks. Diffusion models draw scripts as shapes, so they smear. POCKET-Image renders every glyph exactly — 한국어 · 中文 · 日本語 · العربية (RTL) · ไทย · Latin and more — onto any scene you describe.
What it is:
• 100% accurate text, any language — where global models produce gibberish
• Any background from a prompt — text is optional (empty → a pure image)
• No GPU, no NPU — runs on plain CPU + RAM via the POCKET-Core engine
• Measured footprint: 8.6 GB (RTX 3050/4060) · 4.5 GB (offloaded, 6 GB cards) · 13.4 GB (MacBook, 16 GB+)
• Windows · macOS · Linux · fully local, no cloud
Built on the open, commercial-friendly Z-Image (Apache-2.0) foundation.
Honest note: the text is the guaranteed-correct part — the surrounding scene is ordinary generation, so a busy foreground can crowd the letters. We say so; clean backgrounds stay razor-sharp.
🎨 Studio — generate right here, any language:
FINAL-Bench/POCKET-Image-Studio
🧩 Model card:
FINAL-Bench/POCKET-Image-Zimage
📚 The POCKET collection:
https://huggingface.co/collections/FINAL-Bench/pocket-models
SeaWolf-AI
updated a
collection 9 days ago
SeaWolf-AI
published a
model 9 days ago
SeaWolf-AI
updated a
Space 9 days ago
SeaWolf-AI
published a
Space 9 days ago
SeaWolf-AI
updated 2
Spaces 9 days ago