PaddlePaddle/HPD-Parsing
Image-Text-to-Text โข 1B โข Updated โข 900 โข 67
BuiDoan
BuiDoan
AI & ML interests
None yet
Recent Activity
new activity about 12 hours ago
PaddlePaddle/HPD-Parsing:ๆๆๅฅฝๅทฎ liked a model 2 days ago
PaddlePaddle/HPD-Parsing liked a model 3 days ago
ATH-MaaS/OvisOCR2Organizations
ๆๆๅฅฝๅทฎ
๐ 1
5
#2 opened 5 days ago
by 
AlexLee111
liked a
model 2 days ago
liked a
model 3 days ago
liked 4
models 4 days ago
poolside/Laguna-S-2.1
Text Generation โข 118B โข Updated โข 63.6k โข โข 727
cyankiwi/Hy3-AWQ-INT4
Text Generation โข 50B โข Updated โข 2.54k โข 4
bottlecapai/ThinkingCap-Qwen3.6-27B
Image-Text-to-Text โข 27B โข Updated โข 28.4k โข โข 555
cyankiwi/Laguna-XS-2.1-AWQ-INT4
Text Generation โข 34B โข Updated โข 79 โข 2
reacted to SeaWolf-AI's post with ๐ 4 days ago
Post
3979
๐ฑ POCKET โ 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 โ 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):
โข CPU generate (Xeon, 16 threads): 27.0 vs 10.1 tok/s โ 2.69ร faster
โข GPU generate (H100): 197 vs 89 tok/s โ 2.22ร faster
โข GPU prompt processing (H100): 753 vs 1816 โ 0.41ร (Bonsai wins this one โ MoE prefill wakes every expert, so sparsity stops helping there. We say so.)
โข Quality (HellaSwag, 400 q): 61.0% vs 60.0% โ a tie (confidence intervals overlap)
On a real consumer laptop โ MacBook M3 Pro (18 GB) โ POCKET wins every axis, prompt processing included:
โข Metal generate: 25.4 vs 12.8 โ 1.99ร
โข CPU generate: 13.8 vs 4.4 โ 3.13ร
โข Metal prompt: 240.7 vs 73.4 โ 3.28ร
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 โ it needs the PrismML fork. POCKET runs on the tools you already have: LM Studio, Ollama, PocketPal, MLX.
๐ Full story (tech, measurements, recipes): https://huggingface.co/blog/FINAL-Bench/pocket
Models:
๐ฆ POCKET-35B-GGUF (PC / server, no GPU): FINAL-Bench/POCKET-35B-GGUF
๐ฐ๐ท POCKET-KR-GGUF (Android): FINAL-Bench/POCKET-KR-GGUF
๐ POCKET-KR-MLX (iPhone / Mac): FINAL-Bench/POCKET-KR-MLX
๐ POCKET-EN-GGUF (English phone / PC): FINAL-Bench/POCKET-EN-GGUF
๐ฅ๏ธ Live demo (answering on a CPU, no GPU): FINAL-Bench/POCKET-35B-CPU
๐ 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 โ 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):
โข CPU generate (Xeon, 16 threads): 27.0 vs 10.1 tok/s โ 2.69ร faster
โข GPU generate (H100): 197 vs 89 tok/s โ 2.22ร faster
โข GPU prompt processing (H100): 753 vs 1816 โ 0.41ร (Bonsai wins this one โ MoE prefill wakes every expert, so sparsity stops helping there. We say so.)
โข Quality (HellaSwag, 400 q): 61.0% vs 60.0% โ a tie (confidence intervals overlap)
On a real consumer laptop โ MacBook M3 Pro (18 GB) โ POCKET wins every axis, prompt processing included:
โข Metal generate: 25.4 vs 12.8 โ 1.99ร
โข CPU generate: 13.8 vs 4.4 โ 3.13ร
โข Metal prompt: 240.7 vs 73.4 โ 3.28ร
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 โ it needs the PrismML fork. POCKET runs on the tools you already have: LM Studio, Ollama, PocketPal, MLX.
๐ Full story (tech, measurements, recipes): https://huggingface.co/blog/FINAL-Bench/pocket
Models:
๐ฆ POCKET-35B-GGUF (PC / server, no GPU): FINAL-Bench/POCKET-35B-GGUF
๐ฐ๐ท POCKET-KR-GGUF (Android): FINAL-Bench/POCKET-KR-GGUF
๐ POCKET-KR-MLX (iPhone / Mac): FINAL-Bench/POCKET-KR-MLX
๐ POCKET-EN-GGUF (English phone / PC): FINAL-Bench/POCKET-EN-GGUF
๐ฅ๏ธ Live demo (answering on a CPU, no GPU): FINAL-Bench/POCKET-35B-CPU
๐ Collection: FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6
- 5 replies
liked a
model 9 days ago
upvoted an article 10 days ago
Article
Welcome Inkling by Thinking Machines
- +2
burtenshaw, merve, pcuenq, ariG23498
โข โข 129 liked a
model 11 days ago
liked 3
models 17 days ago
IDEA: Bitnet 1.58 (a4.8) version in future variants would be so incredible!
โค๏ธ 4
3
#10 opened 20 days ago
by 
apiarium
commented on Qwen3.6-27B-FP8 on One RTX 6000 Ada: Fast TTFT, 314 tok/s Decode Generation [Benchmark] 21 days ago
- Quantization: Utilize 4-bit quantization (e.g., AWQ, GPTQ) to optimize GPU memory usage, significantly improving TTFT (Time To First Token) and increasing total throughput (token generation rate).
- Inference Acceleration: Leverage MTP to enhance inference efficiency, achieving a performance boost of 1.3xโ2.0x.
- Context Management: Configure the --max-model-len parameter (typically 32kโ64k) to expand model capability; note that this significantly impacts KV Cache memory footprint, requiring careful VRAM allocation management.
reacted to ginigen-ai's post with โค๏ธ 26 days ago
Post
10469
๐ง Does your LLM know when it's about to be wrong?
Most leaderboards measure accuracy. We measure metacognition โ whether a model catches its own errors. Benchmark + leaderboard + adapters, all open. ๐
The surprise: even a K-AI #1 model (JGOS-31B-Citizen) is the strongest on multiple-choice traps (trap_rate 0.005 โ ~2 misses in 400) yet blind to its own free-form mistakes (self-confidence AUROC = 0.5, pure random). A tiny base-frozen adapter recovers that signal.
Two independent axes (never compared across a row): โ trap_rate โ does it fall for tempting trap options? (lower = stronger) โก adapter gain ฮ โ how much a lightweight adapter catches errors the model itself misses. (higher = more adapter value)
What's open: ๐ 300+100 trap problems (each with a hidden trap + TICOS type) ๐ 24-model leaderboard ๐งฉ 11 per-model adapters โ adapters, NOT fine-tunes (base stays frozen; the adapter just reads the hidden state โ P(wrong))
Submit any HF model โ auto-scored daily at 09:00 KST and added to the board.
๐ Leaderboard โ ginigen-ai/Metacognition-Leaderboard-Space
๐ Benchmark โ ginigen-ai/Metacognition-Bench
๐งฉ Adapters โ FINAL-Bench/metacognition-adapters-6a42c032e6beb803dd032961
๐ Article โ https://huggingface.co/blog/ginigen-ai/metacognition
Benchmark by ginigen-ai ยท Adapters by FINAL-Bench (Darwin/Chimera platform + AETHER metacognition tech).
Most leaderboards measure accuracy. We measure metacognition โ whether a model catches its own errors. Benchmark + leaderboard + adapters, all open. ๐
The surprise: even a K-AI #1 model (JGOS-31B-Citizen) is the strongest on multiple-choice traps (trap_rate 0.005 โ ~2 misses in 400) yet blind to its own free-form mistakes (self-confidence AUROC = 0.5, pure random). A tiny base-frozen adapter recovers that signal.
Two independent axes (never compared across a row): โ trap_rate โ does it fall for tempting trap options? (lower = stronger) โก adapter gain ฮ โ how much a lightweight adapter catches errors the model itself misses. (higher = more adapter value)
What's open: ๐ 300+100 trap problems (each with a hidden trap + TICOS type) ๐ 24-model leaderboard ๐งฉ 11 per-model adapters โ adapters, NOT fine-tunes (base stays frozen; the adapter just reads the hidden state โ P(wrong))
Submit any HF model โ auto-scored daily at 09:00 KST and added to the board.
๐ Leaderboard โ ginigen-ai/Metacognition-Leaderboard-Space
๐ Benchmark โ ginigen-ai/Metacognition-Bench
๐งฉ Adapters โ FINAL-Bench/metacognition-adapters-6a42c032e6beb803dd032961
๐ Article โ https://huggingface.co/blog/ginigen-ai/metacognition
Benchmark by ginigen-ai ยท Adapters by FINAL-Bench (Darwin/Chimera platform + AETHER metacognition tech).
- 11 replies
liked 3
models about 1 month ago
