Quantized variants share exact source weights within each family. Measured QFS KL; RTN format tests, not optimizer-quality rankings.
Michel Belleau PRO
malaiwah
AI & ML interests
None yet
Recent Activity
new activity 11 days ago
deepseek-ai/DeepSeek-V4.1-Flash:Native Transformers text-backbone support proposal (#48721) new activity 12 days ago
deepseek-ai/DeepSeek-V4.1-Flash:Restore each indexer's K cache on incomplete compression stepsOrganizations
None yet
Tiny random models, quantized format fixtures and verified CPU captures for QFS. Pipeline tests, not assistants or quality benchmarks.
-
malaiwah/glm-moe-dsa-tiny-random-bf16
Text Generation • 278k • Updated • 526 -
malaiwah/qwen3-5-tiny-random-bf16
Text Generation • 245k • Updated • 529 -
malaiwah/glm5-next-tiny-random-bf16
Image-Text-to-Text • 422k • Updated • 55 -
malaiwah/qwen4-exp-tiny-random-bf16
Text Generation • 342k • Updated • 498
Four runtime-specific Qwen3.8 EXL3 builds. v5 10M shared-head, body-only KL vs BF16; advisory, not a quality ranking. Context tested separately.
-
malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated
Image-Text-to-Text • 11B • Updated • 612 • 11 -
malaiwah/Qwen3.8-27B-EXL3-K5K6
Image-Text-to-Text • 15B • Updated • 423 • 2 -
malaiwah/Qwen3.8-27B-K4
Image-Text-to-Text • 14B • Updated • 1.09k -
malaiwah/Qwen3.8-27B-EXL3-K5K6-context
Image-Text-to-Text • 10B • Updated • 398 • 8
Quantized variants share exact source weights within each family. Measured QFS KL; RTN format tests, not optimizer-quality rankings.
Independent tiny random native architectures and qualified CPU roots. No trained weights or production-model fine-tuning lineage.
Tiny random models, quantized format fixtures and verified CPU captures for QFS. Pipeline tests, not assistants or quality benchmarks.
-
malaiwah/glm-moe-dsa-tiny-random-bf16
Text Generation • 278k • Updated • 526 -
malaiwah/qwen3-5-tiny-random-bf16
Text Generation • 245k • Updated • 529 -
malaiwah/glm5-next-tiny-random-bf16
Image-Text-to-Text • 422k • Updated • 55 -
malaiwah/qwen4-exp-tiny-random-bf16
Text Generation • 342k • Updated • 498
GLM-5.3-Flash receipt-backed fidelity: named panels, lanes and capture scopes. Descriptive evidence, not universal or native-serving rankings.
Four runtime-specific Qwen3.8 EXL3 builds. v5 10M shared-head, body-only KL vs BF16; advisory, not a quality ranking. Context tested separately.
-
malaiwah/Qwen3.8-27B-EXL3-K5K6-hydrated
Image-Text-to-Text • 11B • Updated • 612 • 11 -
malaiwah/Qwen3.8-27B-EXL3-K5K6
Image-Text-to-Text • 15B • Updated • 423 • 2 -
malaiwah/Qwen3.8-27B-K4
Image-Text-to-Text • 14B • Updated • 1.09k -
malaiwah/Qwen3.8-27B-EXL3-K5K6-context
Image-Text-to-Text • 10B • Updated • 398 • 8