openbmb/MiniCPM5-1B
Text Generation โข 1B โข Updated โข 36.7k โข 652
Rod
Viac
AI & ML interests
playing around with a tool
Recent Activity
liked a model 2 days ago
openbmb/MiniCPM5-1B liked a model 19 days ago
Vikhrmodels/Borealis-5b-it liked a model about 1 month ago
ACE-Step/acestep-v15-xl-baseOrganizations
None yet
liked a
model 2 days ago
liked a
model 19 days ago
liked 4
models about 1 month ago
ACE-Step/acestep-v15-xl-base
Text-to-Audio โข 5B โข Updated โข 1.7k โข 79
Qwen/Qwen3.6-27B
Image-Text-to-Text โข 28B โข Updated โข 5.06M โข โข 1.55k
moonshotai/Kimi-K2.6
Image-Text-to-Text โข 1.1T โข Updated โข 3.07M โข โข 1.38k
lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-IQ4_XS-GGUF
Text Generation โข 35B โข Updated โข 13k โข 29
liked 7
models about 2 months ago
Qwen/Qwen3.6-35B-A3B-FP8
Image-Text-to-Text โข 36B โข Updated โข 5.35M โข 239
nvidia/Lyra-2.0
Image-to-3D โข Updated โข 1.3k โข 326
varjosoft/GLM-4.7-Flash-TQ3
13B โข Updated โข 173 โข 1
Youssofal/MiniMax-M2.7-Abliterated-Heretic-GGUF
Text Generation โข 229B โข Updated โข 2.6k โข 47
MiniMaxAI/MiniMax-M2.7
Text Generation โข 229B โข Updated โข 1.72M โข โข 1.17k
nwzjk/GLM-5.1-Open-TQ3
Text Generation โข 289B โข Updated โข 6 โข 2
ruv/ruvltra-claude-code
Text Generation โข 0.5B โข Updated โข 8.3k โข 181
liked a
dataset about 2 months ago
reacted to reaperdoesntknow's post with ๐ 2 months ago
Post
3287
We present a methodology for training small language models on CPU at FP32 precision
that achieves capability-per-dollar efficiency orders of magnitude beyond GPU-based training.
Across15modelsspanningfournovelarchitecturefamiliesโMixtureofAttentions(MoA),cross-
architecture fusion (Qemma), swarm intelligence (SAGI), and metric-space causal language
models (DiscoverLM)โtotal compute cost was $24 on a single AMD EPYC 9454P proces-
sor. We introduce seven methodological pillars: (1) FP32 precision preservation, with exper-
iments demonstrating 5,810รsingle-operation error and 23,225รcompounding error ratio for
FP16 at network depth; (2) sparse cognitive architectures where 0.02โ7% of parameters activate
per token, matching CPU branching rather than GPU SIMD; (3) developmental curriculum
training progressing from language to logic to transfer to depth; (4) continuous belt-fed data
ingestion eliminating truncation waste; (5) hardware-native optimization for AMD Zen 4 via
AOCL/OpenMP/NUMA-aware allocation; (6) self-regulating thermodynamic governance with
emergent temperature measurement grounded in L2-star discrepancy; and (7) open-standard
compute (AVX2 SIMD at FP32) free of proprietary vendor dependency. We argue that transformers were designed for GPU hardware rather than mathematical optimality, and that architecture designed for geometric correctnessโmetric-space attention, triangle inequality enforcement, sparse expert routingโnaturally favor CPU execution. For sub-2B parameter models, CPU training produces more capable models at a fraction of the cost.
that achieves capability-per-dollar efficiency orders of magnitude beyond GPU-based training.
Across15modelsspanningfournovelarchitecturefamiliesโMixtureofAttentions(MoA),cross-
architecture fusion (Qemma), swarm intelligence (SAGI), and metric-space causal language
models (DiscoverLM)โtotal compute cost was $24 on a single AMD EPYC 9454P proces-
sor. We introduce seven methodological pillars: (1) FP32 precision preservation, with exper-
iments demonstrating 5,810รsingle-operation error and 23,225รcompounding error ratio for
FP16 at network depth; (2) sparse cognitive architectures where 0.02โ7% of parameters activate
per token, matching CPU branching rather than GPU SIMD; (3) developmental curriculum
training progressing from language to logic to transfer to depth; (4) continuous belt-fed data
ingestion eliminating truncation waste; (5) hardware-native optimization for AMD Zen 4 via
AOCL/OpenMP/NUMA-aware allocation; (6) self-regulating thermodynamic governance with
emergent temperature measurement grounded in L2-star discrepancy; and (7) open-standard
compute (AVX2 SIMD at FP32) free of proprietary vendor dependency. We argue that transformers were designed for GPU hardware rather than mathematical optimality, and that architecture designed for geometric correctnessโmetric-space attention, triangle inequality enforcement, sparse expert routingโnaturally favor CPU execution. For sub-2B parameter models, CPU training produces more capable models at a fraction of the cost.
- 6 replies
liked 2
models 2 months ago
reacted to abdurrahmanbutler's post with ๐ค 3 months ago
Post
2591
๐ ๐๐ป๐๐ฟ๐ผ๐ฑ๐๐ฐ๐ถ๐ป๐ด ๐๐ฎ๐ป๐ผ๐ป ๐ฎ ๐๐ป๐ฟ๐ถ๐ฐ๐ต๐ฒ๐ฟ: ๐๐ต๐ฒ ๐๐ผ๐ฟ๐น๐ฑโ๐ ๐ณ๐ถ๐ฟ๐๐ ๐ต๐ถ๐ฒ๐ฟ๐ฎ๐ฟ๐ฐ๐ต๐ถ๐ฐ๐ฎ๐น ๐ด๐ฟ๐ฎ๐ฝ๐ต๐ถ๐๐ถ๐๐ฎ๐๐ถ๐ผ๐ป ๐บ๐ผ๐ฑ๐ฒ๐น
Today weโre publicly releasing Kanon 2 Enricher, and with it, an entirely new class of AI model that weโre calling a hierarchical graphitization model.
This is fundamentally different from both universal extraction models and generative models.
As a hierarchical graphitization model, Kanon 2 Enricher natively outputs a ๐ธ๐ป๐ผ๐๐น๐ฒ๐ฑ๐ด๐ฒ ๐ด๐ฟ๐ฎ๐ฝ๐ต rather than tokens, which makes it architecturally incapable of hallucinating or inventing text that wasnโt present in the input.
What that enables in practice is unlike any other model or ML architecture on the market:
โข ๐ก๐ผ ๐ต๐ฎ๐น๐น๐๐ฐ๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป๐ ๐ค
It cannot hallucinate. All references and links are stored as spans, meaning exact character offsets anchored to the original text.
โข ๐๐ถ๐ฒ๐ฟ๐ฎ๐ฟ๐ฐ๐ต๐ถ๐ฐ๐ฎ๐น ๐๐ฒ๐ด๐บ๐ฒ๐ป๐๐ฎ๐๐ถ๐ผ๐ป, ๐ป๐ผ๐ ๐ท๐๐๐ ๐ฒ๐ ๐๐ฟ๐ฎ๐ฐ๐๐ถ๐ผ๐ป ๐
It deconstructs a documentโs full nested hierarchy, down to chapters, sections, clauses, schedules, signatures, and even singular sentences, and classifies each span with dozens of contextual features.
โข ๐๐ป๐๐ถ๐๐ ๐ฒ๐ ๐๐ฟ๐ฎ๐ฐ๐๐ถ๐ผ๐ป, ๐ฑ๐ถ๐๐ฎ๐บ๐ฏ๐ถ๐ด๐๐ฎ๐๐ถ๐ผ๐ป, ๐ฎ๐ป๐ฑ ๐น๐ถ๐ป๐ธ๐ถ๐ป๐ด ๐
It resolves what references actually point to, then links entities, citations, and cross-references into a single coherent graph.
โข ๐๐ฟ๐ฎ๐ฝ๐ต-๐ณ๐ถ๐ฟ๐๐ ๐ฒ๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฒ๐ป๐ฐ๐ ๐โโก๏ธ
Small enough to run locally on a consumer PC with sub-second latency, and it stays reliable on long documents where front
To read more about our new model, check out our latest Hugging Face article:
https://huggingface.co/blog/isaacus/introducing-kanon-2-enricher
Today weโre publicly releasing Kanon 2 Enricher, and with it, an entirely new class of AI model that weโre calling a hierarchical graphitization model.
This is fundamentally different from both universal extraction models and generative models.
As a hierarchical graphitization model, Kanon 2 Enricher natively outputs a ๐ธ๐ป๐ผ๐๐น๐ฒ๐ฑ๐ด๐ฒ ๐ด๐ฟ๐ฎ๐ฝ๐ต rather than tokens, which makes it architecturally incapable of hallucinating or inventing text that wasnโt present in the input.
What that enables in practice is unlike any other model or ML architecture on the market:
โข ๐ก๐ผ ๐ต๐ฎ๐น๐น๐๐ฐ๐ถ๐ป๐ฎ๐๐ถ๐ผ๐ป๐ ๐ค
It cannot hallucinate. All references and links are stored as spans, meaning exact character offsets anchored to the original text.
โข ๐๐ถ๐ฒ๐ฟ๐ฎ๐ฟ๐ฐ๐ต๐ถ๐ฐ๐ฎ๐น ๐๐ฒ๐ด๐บ๐ฒ๐ป๐๐ฎ๐๐ถ๐ผ๐ป, ๐ป๐ผ๐ ๐ท๐๐๐ ๐ฒ๐ ๐๐ฟ๐ฎ๐ฐ๐๐ถ๐ผ๐ป ๐
It deconstructs a documentโs full nested hierarchy, down to chapters, sections, clauses, schedules, signatures, and even singular sentences, and classifies each span with dozens of contextual features.
โข ๐๐ป๐๐ถ๐๐ ๐ฒ๐ ๐๐ฟ๐ฎ๐ฐ๐๐ถ๐ผ๐ป, ๐ฑ๐ถ๐๐ฎ๐บ๐ฏ๐ถ๐ด๐๐ฎ๐๐ถ๐ผ๐ป, ๐ฎ๐ป๐ฑ ๐น๐ถ๐ป๐ธ๐ถ๐ป๐ด ๐
It resolves what references actually point to, then links entities, citations, and cross-references into a single coherent graph.
โข ๐๐ฟ๐ฎ๐ฝ๐ต-๐ณ๐ถ๐ฟ๐๐ ๐ฒ๐ณ๐ณ๐ถ๐ฐ๐ถ๐ฒ๐ป๐ฐ๐ ๐โโก๏ธ
Small enough to run locally on a consumer PC with sub-second latency, and it stays reliable on long documents where front
To read more about our new model, check out our latest Hugging Face article:
https://huggingface.co/blog/isaacus/introducing-kanon-2-enricher
liked a
model 3 months ago
liked a
Space 4 months ago