webAI-Official / text-generation updated 1 week ago

TwIL-LM3

A 3B reasoning model for formal logic tasks, built from HuggingFaceTB/SmolLM3-3B through LoRA supervised fine-tuning, checkpoint fusion, WiSE-FT weight interpolation, and entropy-weighted GRPO reinforcement learning.

Params
3.1 B
Context
65,536
Downloads 30d
384 K
Likes
90
Licence: other Not gated GGUF · SAFETENSORS 1 languages View on Hugging Face ↗

Download history

daily snapshots · 25 days
408 K179 K
Aug 28Sep 5Sep 13Sep 21

Can you run it?

Estimated VRAM at 8K context unless noted. Pick your hardware to see the verdict per quantization.

FileQuantSizeEst. VRAMVerdict on RTX 4090 · 24 GB
TwIL-LM3-Q4_K_M.gguf Q4_K_M 1.9 GB 3.1 GB ✅ Runs comfortably
TwIL-LM3-Q5_K_M.gguf Q5_K_M 2.2 GB 3.4 GB ✅ Runs comfortably
TwIL-LM3-Q6_K.gguf Q6_K 2.5 GB 3.7 GB ✅ Runs comfortably
TwIL-LM3-Q8_0.gguf Q8_0 3.3 GB 4.6 GB ✅ Runs comfortably
TwIL-LM3-F16.gguf GGUF 6.2 GB 7.7 GB ✅ Runs comfortably
model.safetensors bf16 6.2 GB 7.7 GB ✅ Runs comfortably
model.safetensors (bf16, full) bf16 + 64K ctx 6.2 GB 11.0 GB ✅ Runs comfortably
Estimate: file size × 1.1 + KV cache at 8K + 0.5 GB overhead. Not a benchmark — how we calculate this.

Run it

copy-paste, exact tags checked against the Hub
~ · ollama · Q4_K_M
$ ollama run twil-lm3

# pin the quantization explicitly
$ ollama run twil-lm3-q4_k_m
est. VRAM 3.1 GBon RTX 4090 · 24 GBJSON API →

Specifications

Architecture
SmolLM3ForCausalLM
Parameters
3.1 B
Tensor type
BF16
Context length
65,536
Vocabulary
128,256
Layers / heads
36 / 16
Licence
other
First seen on the Hub
2026-08-03
Base model
SmolLM3-3B
Training datasets
undisclosed
Added to our catalog
2026-08-28

Family

Base model and the most-downloaded derivatives in the catalog.

Compare with any text-generation model