Article
The Fast Gemma Challenge: our verified-SOTA recipe, in full
FINAL-Bench
• • 14
YoungMin
Rickie77
AI & ML interests
None yet
Recent Activity
upvoted an article about 5 hours ago
The Fast Gemma Challenge: our verified-SOTA recipe, in full reacted to SeaWolf-AI's post with ❤️ about 5 hours ago
We wrote up our run in The Fast Gemma Challenge — as vidraft-darwin — and wanted to share the recipe. 🙏
https://huggingface.co/spaces/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
reacted to SeaWolf-AI's post with 🚀 about 5 hours ago
We wrote up our run in The Fast Gemma Challenge — as vidraft-darwin — and wanted to share the recipe. 🙏
https://huggingface.co/spaces/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
Organizations
None yet
upvoted an article about 5 hours ago
reacted to SeaWolf-AI's post with ❤️🚀🔥 about 5 hours ago
Post
661
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
reacted to SeaWolf-AI's post with ❤️🔥👍 9 days ago
Post
3120
POCKET now speaks Gemma 4 — a 26B model that loads in every app, and runs on your PC with no GPU
We're adding a Gemma-4 sibling to POCKET: POCKET-26B, built from Google's Gemma-4-26B-A4B (Apache-2.0). Our flagship POCKET-35B is a Qwen-family MoE and needs a recent llama.cpp; POCKET-26B trades a little size for the thing people kept asking for — it just loads, everywhere, today: Ollama, LM Studio, PocketPal, MLX, any stock llama.cpp. No fork, no bleeding-edge runtime, no CUDA, no cloud.
It's a sparse Mixture-of-Experts (25.2B total, ~4B active per token), so the work per token stays small — a real 26B that generates on a CPU with no graphics card.
Two things make it stand out:
1) Universal compatibility. Gemma 4 is a standard, widely-supported architecture, so POCKET-26B runs on the tools you already have — no waiting for your app to add a new model type.
2) Quality that survives compression. Measured GPQA-Diamond (198 q, greedy):
• Full base: 67.7%
• POCKET-26B Q4_K_M (17 GB): 67.7% — lossless
• POCKET-26B Q2_K (11 GB): 67.2% — near-lossless, at 11 GB
Live, on a CPU-only box (our demo Space — POCKET-26B vs Bonsai-27B, same machine, same stock llama.cpp): POCKET-26B ≈ 19 tok/s vs Bonsai ≈ 6 tok/s → about 3× faster generation, no GPU. (Honest notes: shared CPU box, sequential race; a dedicated machine is faster.)
Where it fits in the family:
• POCKET-35B (Qwen MoE) — bigger, top-tier, needs a recent llama.cpp.
• POCKET-26B (Gemma 4) — loads in any app, quality-robust when compressed. The demo runs the Q4_K_M build; Q2_K (11 GB) is the smallest footprint. For a true ≤8 GB phone, the 5 GB POCKET-KR (Qwen) is still the pick.
Try it and grab it:
🖥️ Live demo (Gemma4-based, answering on a CPU, no GPU): FINAL-Bench/POCKET-26B-CPU
📦 POCKET-26B-GGUF (Q4_K_M 17 GB · Q2_K 11 GB): FINAL-Bench/POCKET-26B-GGUF
📚 POCKET collection: https://huggingface.co/collections/FINAL-Bench/pocket-models
We're adding a Gemma-4 sibling to POCKET: POCKET-26B, built from Google's Gemma-4-26B-A4B (Apache-2.0). Our flagship POCKET-35B is a Qwen-family MoE and needs a recent llama.cpp; POCKET-26B trades a little size for the thing people kept asking for — it just loads, everywhere, today: Ollama, LM Studio, PocketPal, MLX, any stock llama.cpp. No fork, no bleeding-edge runtime, no CUDA, no cloud.
It's a sparse Mixture-of-Experts (25.2B total, ~4B active per token), so the work per token stays small — a real 26B that generates on a CPU with no graphics card.
Two things make it stand out:
1) Universal compatibility. Gemma 4 is a standard, widely-supported architecture, so POCKET-26B runs on the tools you already have — no waiting for your app to add a new model type.
2) Quality that survives compression. Measured GPQA-Diamond (198 q, greedy):
• Full base: 67.7%
• POCKET-26B Q4_K_M (17 GB): 67.7% — lossless
• POCKET-26B Q2_K (11 GB): 67.2% — near-lossless, at 11 GB
Live, on a CPU-only box (our demo Space — POCKET-26B vs Bonsai-27B, same machine, same stock llama.cpp): POCKET-26B ≈ 19 tok/s vs Bonsai ≈ 6 tok/s → about 3× faster generation, no GPU. (Honest notes: shared CPU box, sequential race; a dedicated machine is faster.)
Where it fits in the family:
• POCKET-35B (Qwen MoE) — bigger, top-tier, needs a recent llama.cpp.
• POCKET-26B (Gemma 4) — loads in any app, quality-robust when compressed. The demo runs the Q4_K_M build; Q2_K (11 GB) is the smallest footprint. For a true ≤8 GB phone, the 5 GB POCKET-KR (Qwen) is still the pick.
Try it and grab it:
🖥️ Live demo (Gemma4-based, answering on a CPU, no GPU): FINAL-Bench/POCKET-26B-CPU
📦 POCKET-26B-GGUF (Q4_K_M 17 GB · Q2_K 11 GB): FINAL-Bench/POCKET-26B-GGUF
📚 POCKET collection: https://huggingface.co/collections/FINAL-Bench/pocket-models
- 3 replies
liked a
Space 9 days ago
liked a
model 9 days ago
liked a
Space 11 days ago
reacted to SeaWolf-AI's post with ❤️🤗👍 11 days ago
Post
4028
📱 POCKET — a 35-billion-parameter model that runs on your iPhone, and on your PC with no GPU
We're releasing POCKET, VIDRAFT's flagship Darwin-36B-Opus compressed for on-device use. No fork, no CUDA, no cloud — it runs on stock llama.cpp. It's a sparse Mixture-of-Experts model (256 experts, only 8 active per token), so the file can be large while the work per token stays small. That's what lets a 35B model run on a phone, and generate fast on a CPU with no graphics card.
Measured (POCKET-35B IQ1_M vs Bonsai-27B Q1_0):
• CPU generate (Xeon, 16 threads): 27.0 vs 10.1 tok/s → 2.69× faster
• GPU generate (H100): 197 vs 89 tok/s → 2.22× faster
• GPU prompt processing (H100): 753 vs 1816 → 0.41× (Bonsai wins this one — MoE prefill wakes every expert, so sparsity stops helping there. We say so.)
• Quality (HellaSwag, 400 q): 61.0% vs 60.0% → a tie (confidence intervals overlap)
On a real consumer laptop — MacBook M3 Pro (18 GB) — POCKET wins every axis, prompt processing included:
• Metal generate: 25.4 vs 12.8 → 1.99×
• CPU generate: 13.8 vs 4.4 → 3.13×
• Metal prompt: 240.7 vs 73.4 → 3.28×
One more quiet fact: the same-size, quality-oriented rival Ternary-Bonsai-27B (7.2 GB) fails to load in upstream llama.cpp at all — it needs the PrismML fork. POCKET runs on the tools you already have: LM Studio, Ollama, PocketPal, MLX.
📖 Full story (tech, measurements, recipes): https://huggingface.co/blog/FINAL-Bench/pocket
Models:
📦 POCKET-35B-GGUF (PC / server, no GPU): FINAL-Bench/POCKET-35B-GGUF
🇰🇷 POCKET-KR-GGUF (Android): FINAL-Bench/POCKET-KR-GGUF
🍎 POCKET-KR-MLX (iPhone / Mac): FINAL-Bench/POCKET-KR-MLX
🌍 POCKET-EN-GGUF (English phone / PC): FINAL-Bench/POCKET-EN-GGUF
🖥️ Live demo (answering on a CPU, no GPU): FINAL-Bench/POCKET-35B-CPU
📚 Collection: FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6
We're releasing POCKET, VIDRAFT's flagship Darwin-36B-Opus compressed for on-device use. No fork, no CUDA, no cloud — it runs on stock llama.cpp. It's a sparse Mixture-of-Experts model (256 experts, only 8 active per token), so the file can be large while the work per token stays small. That's what lets a 35B model run on a phone, and generate fast on a CPU with no graphics card.
Measured (POCKET-35B IQ1_M vs Bonsai-27B Q1_0):
• CPU generate (Xeon, 16 threads): 27.0 vs 10.1 tok/s → 2.69× faster
• GPU generate (H100): 197 vs 89 tok/s → 2.22× faster
• GPU prompt processing (H100): 753 vs 1816 → 0.41× (Bonsai wins this one — MoE prefill wakes every expert, so sparsity stops helping there. We say so.)
• Quality (HellaSwag, 400 q): 61.0% vs 60.0% → a tie (confidence intervals overlap)
On a real consumer laptop — MacBook M3 Pro (18 GB) — POCKET wins every axis, prompt processing included:
• Metal generate: 25.4 vs 12.8 → 1.99×
• CPU generate: 13.8 vs 4.4 → 3.13×
• Metal prompt: 240.7 vs 73.4 → 3.28×
One more quiet fact: the same-size, quality-oriented rival Ternary-Bonsai-27B (7.2 GB) fails to load in upstream llama.cpp at all — it needs the PrismML fork. POCKET runs on the tools you already have: LM Studio, Ollama, PocketPal, MLX.
📖 Full story (tech, measurements, recipes): https://huggingface.co/blog/FINAL-Bench/pocket
Models:
📦 POCKET-35B-GGUF (PC / server, no GPU): FINAL-Bench/POCKET-35B-GGUF
🇰🇷 POCKET-KR-GGUF (Android): FINAL-Bench/POCKET-KR-GGUF
🍎 POCKET-KR-MLX (iPhone / Mac): FINAL-Bench/POCKET-KR-MLX
🌍 POCKET-EN-GGUF (English phone / PC): FINAL-Bench/POCKET-EN-GGUF
🖥️ Live demo (answering on a CPU, no GPU): FINAL-Bench/POCKET-35B-CPU
📚 Collection: FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6
- 5 replies
reacted to SeaWolf-AI's post with 🚀🔥❤️ 15 days ago
Post
5143
A small gift for anyone building or studying foundation models.
Most "open" models hand you the weights and stop there. With Aether-7B-5Attn we wanted to hand over the whole thing — so you can actually learn from it, reproduce it, and build on it: the data recipe, the training code, every hyperparameter, the complete logs, and the intermediate checkpoints. All Apache-2.0, reproducible byte-for-byte.
What you can do with it:
🔁 Rebuild it from scratch, or fork the recipe for your own model
🔬 Study a real heterogeneous-attention MoE — 49 layers place 5 attention mechanisms on a 7×7 Latin square, arranged as a clean, attributable ablation
📈 Trace training dynamics across the released checkpoints (110k / 115k / 162k)
It's a modest 6.59B model, and an honest one — the limitations (no KV-cache in this build, small scale) are written right in the card. We're not claiming it's special. If any piece of it saves you time or teaches you something, that's exactly what we hoped for. 🤗
📖 Full write-up →
[blog] · https://huggingface.co/blog/FINAL-Bench/opensource-llm
📦 5 Attention Base · FINAL-Bench/Aether-7B-5Attn
🎯 5 Attention Instruct · FINAL-Bench/Aether-7B-5Attn-it
🚀 5 Attention Live demo · FINAL-Bench/Aether-Sovereign-AI
📦 7 Attention Base · https://huggingface.co/FINAL-Bench/Aether-7B-7Attn-base
📦 11 Attention Base · FINAL-Bench/Aether-6B-11Attn-base
🧬 Collection · https://huggingface.co/collections/FINAL-Bench/aether-foundation-model
#opensource #LLM #MoE #reproducibility #Apache2
Most "open" models hand you the weights and stop there. With Aether-7B-5Attn we wanted to hand over the whole thing — so you can actually learn from it, reproduce it, and build on it: the data recipe, the training code, every hyperparameter, the complete logs, and the intermediate checkpoints. All Apache-2.0, reproducible byte-for-byte.
What you can do with it:
🔁 Rebuild it from scratch, or fork the recipe for your own model
🔬 Study a real heterogeneous-attention MoE — 49 layers place 5 attention mechanisms on a 7×7 Latin square, arranged as a clean, attributable ablation
📈 Trace training dynamics across the released checkpoints (110k / 115k / 162k)
It's a modest 6.59B model, and an honest one — the limitations (no KV-cache in this build, small scale) are written right in the card. We're not claiming it's special. If any piece of it saves you time or teaches you something, that's exactly what we hoped for. 🤗
📖 Full write-up →
[blog] · https://huggingface.co/blog/FINAL-Bench/opensource-llm
📦 5 Attention Base · FINAL-Bench/Aether-7B-5Attn
🎯 5 Attention Instruct · FINAL-Bench/Aether-7B-5Attn-it
🚀 5 Attention Live demo · FINAL-Bench/Aether-Sovereign-AI
📦 7 Attention Base · https://huggingface.co/FINAL-Bench/Aether-7B-7Attn-base
📦 11 Attention Base · FINAL-Bench/Aether-6B-11Attn-base
🧬 Collection · https://huggingface.co/collections/FINAL-Bench/aether-foundation-model
#opensource #LLM #MoE #reproducibility #Apache2
- 5 replies
liked a
dataset 15 days ago
liked a
model 15 days ago
liked a
Space 15 days ago
upvoted a collection 15 days ago