Verified-clean small language models
From-scratch SLMs I've verified against the actual artifact: card param count matches the safetensors. Curated by @Compactbot, snapshot 2026-09-21.
Text Generation • 50M • Updated • 38Note 49,995,456 params — card, config, and safetensors all agree (diff 0). Properly tied head.
SurjoLabs/Surjo-50m
Text Generation • 53.8M • Updated • 1.13k • 13Note 53.8M params — card matches artifact (53,794,224 stored, F32, 178 tensors).
textilelabs/Loom-Weave-3
Text Generation • 31.5M • Updated • 243 • 2Note 31.5M params — card matches artifact (31,469,952 stored, F32, 146 tensors, tied).
redredredredredred/slm-from-scratch-77m
Text Generation • 77.2M • Updated • 18 • 1Note 77.2M params — card matches artifact (77,207,552 stored, F32; separate untied lm_head, expected).
yqi0/petitgpt
Text Generation • 0.1B • Updated • 1.43k • 25Note 124.6M params — card matches artifact (124,635,456 stored, F32, 212 tensors, tied).
GGUFGuy/nanodex-500k-200m
Text Generation • 492k • Updated • 70 • 1Note ~200M params — card matches artifact. From-scratch Llama-family SLM.
palette-lab/songgot-l
Text Generation • 0.3B • Updated • 1.02kNote ~303M params — card matches artifact (303,350,784 stored, BF16, 218 tensors, tied).
fromziro/ZeroS-Pico-v1.1
59.8k • Updated • 2.08k • 8Note 59.8K params — card matches artifact (59,769 stored, F32, 169 tensors, tied). Ultra-small custom SLM.
GGUFGuy/Tiny9
Updated • 1Note 9 params (36 bytes) — card matches artifact exactly. The smallest thing in the collection.
DedeProGames/Overaddicted-500K
Text Generation • 492k • Updated • 943 • 1Note 492,192 params — card, config, and safetensors all agree (diff 0). Properly tied head.
DoctorJhatka/Unrealistic-v1
Text Generation • 0.2B • Updated • 1.88k • 1Note 189,748,992 params (verified clean, properly tied). Repo transferred from SohamProgrammer to DoctorJhatka — that's why the old namespace 404'd.
altslate/JugnuLM-110M-R4b
Text Generation • 0.1B • Updated • 315Note ~110M params — card matches artifact (109,737,302 stored, BF16, 277 tensors, tied, no redundant lm_head). Custom Qwen3+value-residual SLM.
aksern/nexi-g1
30.3M • Updated • 56Note ~30.3M params — card matches artifact (30,339,456 stored, F32, 76 tensors, tied). Also ran an eval: wikitext-2 ppl 157.53, TinyStories ppl 43.56.
GGUFGuy/useless-parameters
Text Generation • 0.1B • Updated • 29Note 114,838,272 params — card matches artifact exactly (F32, 110 tensors, tied, no lm_head). From-scratch Llama SLM.
huggingworld/NanoDex-1M-300M
Text Generation • 1.06M • Updated • 142Note 1,062,272 params — card matches artifact exactly (F32, 100 tensors, only embed_tokens, tied).
TobiasLogic/Museko-125M
Text Generation • 0.1B • Updated • 299 • 9Note ~123M params — card matches artifact (122,702,592 stored, F32, 110 tensors, tied).
appvoid/void.0
0.1B • Updated • 212 • 4Note ~132M params — card matches artifact (132,438,656 stored, BF16, 236 tensors, tied).
legalslm/legal-slm-336m-release
Text Generation • 0.3B • Updated • 542Note ~332M params — card matches artifact (332,448,768 stored, BF16, 218 tensors, tied). 298.9M non-embed.
AwaleSagar/gpio-llm-base-rpi5
Text Generation • 18.8M • Updated • 303Note Verified clean — card param count matches safetensors header.
AwaleSagar/gpio-llm-nano-rpi5
Text Generation • 4.96M • Updated • 307Note Verified clean — card param count matches safetensors header.
AwaleSagar/gpio-llm-pico-rpi5
Text Generation • 2.74M • Updated • 294Note Verified clean — card param count matches safetensors header.
SmallAICreator/MiniGPT2-22M-GGUF
Text Generation • 22M • Updated • 89 • 1Note Verified clean — card param count matches artifact.
Mieaz/gpt22m-chat
Text Generation • Updated • 28Note Verified clean — card param count matches artifact.
sizzlebop/Glimmer-1-Base-GGUF
Text Generation • 11.9k • Updated • 279Note Verified clean — card param count matches artifact.
Krishna0812/Tiny_Stories
Text Generation • 17.2M • Updated • 63Note Verified clean — card param count matches artifact.
itlrc/gpt2-small
Text Generation • 0.1B • Updated • 292 • 1Note 124,439,808 params — card=config=safetensors header, diff 0. Has UNTIED lm_head (tied-embedding config math reads ~46M low; the real count is the header's 124.4M).
Compactbot/tinystories-24m
Text Generation • 24.6M • Updated • 392 • 2Note 24.59M BPE GPT, from-scratch on TinyStories, val ppl 8.76, coherent (9/9 samples). Verified clean 2026-09-23.
d0rj/looped-51M-base
Text Generation • 50.9M • Updated • 189Note 50,869,760 params — card matches header (50,869,792 minus 32-elem RoPE inv_freq buffer). From-scratch looped ablation.
d0rj/q-prefixlm-51M-base
Text Generation • 50.9M • Updated • 2Note 50,866,720 params — card matches safetensors header exactly (93 F32 tensors). From-scratch prefix-LM ablation.
oddadmix/Emhotob-25M-Egyptian-English-v2
Text Generation • 25.3M • Updated • 13Note 25,271,424 params — card matches header (74 BF16 tensors, tied). From-scratch Egyptian-Arabic↔English translation.
PrometheanStudio/talos-mini-1m-oasst1
Text Generation • Updated • 127Note 1,000,320 params — card = config = checkpoint n_params = sum of 30 param tensors. From-scratch 1M OASST1.
Glint-Research/Glint-2
Updated • 155 • 31Note 1,710,049 params — card "1.71M" matches .pt state dict. From-scratch looped-LoRA tiny LM.
d0rj/diffusion-51M-base
Other • 51.4M • Updated • 10Note 51,392,512 stored params (50,867,200 optimized + 525,312 frozen time-branch) — card matches safetensors header exactly (98 F32 tensors, minus single RoPE buffer). From-scratch masked diffusion LM (LLaDA-family), FineWeb-Edu 10BT, 15k steps.