---

license: apache-2.0

base_model: SupraLabs/Supra2-100M-Instruct

tags:

- gguf

- llama.cpp

- qwen3

- chatml

quantized_by: Felladrin

---

gguf-f16-Supra2-100M-Instruct

GGUF conversion of SupraLabs/Supra2-100M-Instruct, produced with llama.cpp's convert_hf_to_gguf.py.

• Supra2-100M-Instruct-F16.gguf: full-precision (F16) GGUF, unquantized, with the ChatML template embedded.

One thing here differs from a faithful conversion: the source declares </s> as its EOS token, while its chat template ends every turn with <|im_end|>. In this file <|im_end|> (id 6) is registered as both EOS and EOT, so a runtime stops at the end of the turn without needing extra stop strings. llama.cpp already stops there by itself (it treats the <|im_end|> name as end-of-generation, whatever the EOS metadata says), so this only matters for runtimes without that heuristic.

Everything else is untouched: all 134 tensors are byte-identical (sha256 per tensor) to the Supra2-100M-SFT-F16.gguf that SupraLabs publish in the source repo. If you run llama.cpp, either file gives you the same model.

Note: llama.cpp doesn't recognize this model's BPE pre-tokenizer, so the conversion sets it to qwen2, which does the same thing the source tokenizer does (Digits with individual digits, then ByteLevel). That's the same mapping the source GGUF uses.

See the source model card for training details, license terms and usage notes.