PollardWeights/Qwen3-ASR-1.7B-Pollard
Automatic Speech Recognition • 2B • Updated • 3.45k • 2
AI & ML interests
Memory-fit quantization · role-aware GGUF builds · running frontier LLMs on small hardware · MoE expert placement
Recent Activity
View all activity
Organization Card
Frontier models. Small hardware. No compromise.
Pollard Weights builds a model measured to fit your target device — phone to wafer — and ships it to any runtime: GGUF · MLX · GPTQ · MX · EXL3 · EXLR8 · DRAFT.
- 🔧 The tool: github.com/WestWaters/pollard-weights
- 📦 This org hosts Pollard-built model repos (know-what-fits-before-you-download).
- One measured allocation, seven lanes:
| Lane | What ships | Runs on |
|---|---|---|
| GGUF | full ladder (Q6_K · IQ4_XS · IQ3_S · 2-bit) + the IQ*_KT trellis flagship, mmproj for vision/audio | llama.cpp / Ollama / LM Studio (stock); flagship on ik_llama |
| MLX | 2 / 3 / 4 / 5 / 6 / 8-bit, measured allocation | Apple Silicon (mlx-lm / mlx-vlm) |
| GPTQ | int2–int8, dynamic per-layer 4/8 mix | vLLM / SGLang |
| MX | NVFP4 · MXFP4 · FP8 · MXFP6 E2M3 / E3M2 · measured FP8/MXFP6/FP4 mixes | Blackwell / Rubin via vLLM (FP8 also Hopper / Ada) |
| EXL3 | several bpw, smoothed + Calib 3.0 | exllamav3 |
| EXLR8 | one trellis encode, every bpw a manifest (2.5–6 bpw) | GR8 / reference decode |
| DRAFT | speculative-decoding drafter (DFlash / DSpark / EAGLE-3) measured against the quantized target | llama.cpp / vLLM / SGLang / MLX |
- 🧩 Every lane ships the whole model — vision and audio towers, projectors, processors. Never silently text-only.
- 🛠️ Tool-call aware on every build: calibration, gate and card.
- 📁 One repo per model, every lane inside: GGUF at the root,
MLX/ GPTQ/ MX/ EXL3/ EXLR8/ DRAFT/folders. - 📊 Measured numbers on every card: KLD, top-1, PPL and tool calls per rung.
models 34
PollardWeights/Ming-3.0-tts-0.5B-Pollard
Text-to-Speech • Updated • 291
PollardWeights/LFM3-Audio-2B-Pollard
Any-to-Any • 2B • Updated • 1.6k • 1
PollardWeights/LFM3-Omni-2B-Pollard
Any-to-Any • 2B • Updated • 3.2k • 2
PollardWeights/LFM3-1.7B-Pollard
Text Generation • 2B • Updated • 1.04k
PollardWeights/granite-speech-5.0-2b-Pollard
Automatic Speech Recognition • 2B • Updated • 445
PollardWeights/Nanbeige4.3-33B-Pollard
Text Generation • 33B • Updated • 1.68k • 1
PollardWeights/NVIDIA-Nemotron-3.5-Bolt-E4B-Pollard
Text Generation • 8B • Updated • 8.36k • 2
PollardWeights/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Pollard
Text Generation • 74B • Updated • 3.12k • 1
PollardWeights/clef-omni-Pollard
Text Classification • 31B • Updated • 1.59k
datasets 0
None public yet