An open world model for agents. Paper: doi.org/10.5281/zenodo.23146970
caio vicentino PRO
caiovicentino1
AI & ML interests
None yet
Recent Activity
updated a dataset about 16 hours ago
caiovicentino1/ekbasis-data updated a collection about 16 hours ago
Ekbasis updated a collection about 16 hours ago
EkbasisOrganizations
Agent-safety interpretability arc, first public Qwen3.6 SAEs, the probe/guard suite, and the agent-trajectory benchmarks behind them. openinterp.org
Hadamard-Lloyd Weight Quantization · arXiv:2603.29078 · formerly PolarQuant
-
caiovicentino1/Qwen3.5-9B-HLWQ-Q5
Text Generation • 9B • Updated • 39 • 3 -
caiovicentino1/Qwen3.5-9B-HLWQ-MLX-4bit
Text Generation • 9B • Updated • 70 • 5 -
caiovicentino1/Qwen3.5-27B-HLWQ-Q5
Text Generation • 27B • Updated • 77 • 10 -
caiovicentino1/Qwen3.5-9B-HLWQ-Engine-v4
Text Generation • 7B • Updated • 21
Google Gemma family quantized with HLWQ (Hadamard-Lloyd) · formerly PolarQuant Gemma
-
caiovicentino1/Gemma-4-31B-it-HLWQ-Q5
Text Generation • Updated • 26 • 4 -
caiovicentino1/Gemma-4-31B-it-HLWQ-Q5-Vision
Image-Text-to-Text • Updated • 30 • 7 -
caiovicentino1/Gemma-4-26B-A4B-it-HLWQ-Q5
Image-Text-to-Text • 27B • Updated • 29 • 8 -
caiovicentino1/Gemma-4-31B-Claude-Opus-HLWQ-Q5-Vision
Image-Text-to-Text • Updated • 90 • 18
Full-stack HLWQ: Q5 weights + torchao INT4 + Q3 KV cache · formerly PolarQuant Unified
-
caiovicentino1/Qwopus3.5-9B-v3-HLWQ-Q5
Text Generation • 3B • Updated • 43 • 9 -
caiovicentino1/Qwen3.5-9B-Claude-Opus-HLWQ-Q5
Text Generation • 9B • Updated • 91 • 5 -
caiovicentino1/Qwen3.5-27B-Claude-Opus-HLWQ-Q5
Text Generation • 27B • Updated • 14 -
caiovicentino1/Qwopus3.5-9B-v3-HLWQ-MLX-4bit
Text Generation • 9B • Updated • 35 • 8
HLWQ + EOQ quantized large models · Claude Opus distilled + MoE variants
-
caiovicentino1/Qwen3.5-27B-HLWQ-Q5
Text Generation • 27B • Updated • 77 • 10 -
caiovicentino1/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-EOQ-Q5-compressed
27B • Updated • 10 • 1 -
caiovicentino1/Qwen3.5-35B-A3B-EOQ-v3
15B • Updated • 22 -
caiovicentino1/Qwen3.5-35B-A3B-EOQ-Q5-compressed
35B • Updated • 15 • 1
EOQ quantized Qwen2.5 models (Q4/Q5/Q6/Q8). Dequant at load, zero inference overhead.
EOQ (Entropy-Optimal Quantization) compressed models. Mixed-bit allocation + rANS entropy coding. Smaller download, dequant at load time.
Open, calibrated, single-pass typed-decision models (4B & 27B) for finance and trading
Massive MoE models ≥100B quantized with HLWQ · consumer deploy via vLLM expert offload
Video & diffusion models quantized with HLWQ Q5 · 50-65% smaller · formerly PolarQuant
30B MoE · 7.6 GB VRAM · 15 tok/s on RTX 4090 · expert offloading + HLWQ Q5
HLWQ models for Apple Silicon via MLX · run LLMs on Mac · formerly PolarQuant MLX
EOQ quantized Qwen3.5-4B models (Q4/Q5/Q6).
Qwen3.5-9B · HLWQ Q5 · beats torchao INT4 on PPL (6.56 vs 6.68) · CUDA + MLX
-
caiovicentino1/Qwen3.5-9B-HLWQ-Q5
Text Generation • 9B • Updated • 39 • 3 -
caiovicentino1/Qwen3.5-9B-HLWQ-MLX-4bit
Text Generation • 9B • Updated • 70 • 5 -
caiovicentino1/Qwen3.5-9B-HLWQ-Engine-v4
Text Generation • 7B • Updated • 21 -
caiovicentino1/Qwen3.5-9B-EOQ-v3
Text Generation • 5B • Updated • 20 • 2
Qwen3.5-27B · HLWQ Q5 weight quantization · formerly PolarQuant
An open world model for agents. Paper: doi.org/10.5281/zenodo.23146970
Open, calibrated, single-pass typed-decision models (4B & 27B) for finance and trading
Agent-safety interpretability arc, first public Qwen3.6 SAEs, the probe/guard suite, and the agent-trajectory benchmarks behind them. openinterp.org
Massive MoE models ≥100B quantized with HLWQ · consumer deploy via vLLM expert offload
Hadamard-Lloyd Weight Quantization · arXiv:2603.29078 · formerly PolarQuant
-
caiovicentino1/Qwen3.5-9B-HLWQ-Q5
Text Generation • 9B • Updated • 39 • 3 -
caiovicentino1/Qwen3.5-9B-HLWQ-MLX-4bit
Text Generation • 9B • Updated • 70 • 5 -
caiovicentino1/Qwen3.5-27B-HLWQ-Q5
Text Generation • 27B • Updated • 77 • 10 -
caiovicentino1/Qwen3.5-9B-HLWQ-Engine-v4
Text Generation • 7B • Updated • 21
Video & diffusion models quantized with HLWQ Q5 · 50-65% smaller · formerly PolarQuant
Google Gemma family quantized with HLWQ (Hadamard-Lloyd) · formerly PolarQuant Gemma
-
caiovicentino1/Gemma-4-31B-it-HLWQ-Q5
Text Generation • Updated • 26 • 4 -
caiovicentino1/Gemma-4-31B-it-HLWQ-Q5-Vision
Image-Text-to-Text • Updated • 30 • 7 -
caiovicentino1/Gemma-4-26B-A4B-it-HLWQ-Q5
Image-Text-to-Text • 27B • Updated • 29 • 8 -
caiovicentino1/Gemma-4-31B-Claude-Opus-HLWQ-Q5-Vision
Image-Text-to-Text • Updated • 90 • 18
30B MoE · 7.6 GB VRAM · 15 tok/s on RTX 4090 · expert offloading + HLWQ Q5
Full-stack HLWQ: Q5 weights + torchao INT4 + Q3 KV cache · formerly PolarQuant Unified
-
caiovicentino1/Qwopus3.5-9B-v3-HLWQ-Q5
Text Generation • 3B • Updated • 43 • 9 -
caiovicentino1/Qwen3.5-9B-Claude-Opus-HLWQ-Q5
Text Generation • 9B • Updated • 91 • 5 -
caiovicentino1/Qwen3.5-27B-Claude-Opus-HLWQ-Q5
Text Generation • 27B • Updated • 14 -
caiovicentino1/Qwopus3.5-9B-v3-HLWQ-MLX-4bit
Text Generation • 9B • Updated • 35 • 8
HLWQ models for Apple Silicon via MLX · run LLMs on Mac · formerly PolarQuant MLX
HLWQ + EOQ quantized large models · Claude Opus distilled + MoE variants
-
caiovicentino1/Qwen3.5-27B-HLWQ-Q5
Text Generation • 27B • Updated • 77 • 10 -
caiovicentino1/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-EOQ-Q5-compressed
27B • Updated • 10 • 1 -
caiovicentino1/Qwen3.5-35B-A3B-EOQ-v3
15B • Updated • 22 -
caiovicentino1/Qwen3.5-35B-A3B-EOQ-Q5-compressed
35B • Updated • 15 • 1
EOQ quantized Qwen3.5-4B models (Q4/Q5/Q6).
EOQ quantized Qwen2.5 models (Q4/Q5/Q6/Q8). Dequant at load, zero inference overhead.
Qwen3.5-9B · HLWQ Q5 · beats torchao INT4 on PPL (6.56 vs 6.68) · CUDA + MLX
-
caiovicentino1/Qwen3.5-9B-HLWQ-Q5
Text Generation • 9B • Updated • 39 • 3 -
caiovicentino1/Qwen3.5-9B-HLWQ-MLX-4bit
Text Generation • 9B • Updated • 70 • 5 -
caiovicentino1/Qwen3.5-9B-HLWQ-Engine-v4
Text Generation • 7B • Updated • 21 -
caiovicentino1/Qwen3.5-9B-EOQ-v3
Text Generation • 5B • Updated • 20 • 2
EOQ (Entropy-Optimal Quantization) compressed models. Mixed-bit allocation + rANS entropy coding. Smaller download, dequant at load time.
Qwen3.5-27B · HLWQ Q5 weight quantization · formerly PolarQuant