Architecture: DFlashLagunaForCausalLM (6 sliding-attention layers, block_size 16)
Quantization: 4.00 bpw (head: 6.00 bpw), single 535 MB shard
Shares token embedding + lm_head with the target; draft_vocab_size == vocab_size
Verified: config.json + quantization_config.json present, conversion completed cleanly

[!IMPORTANT] This is a draft/speculator model, not a standalone chat model. It cannot generate text on its own — it proposes candidate tokens that the target model verifies. Use it as the draft model in a speculative decoding config, paired with a quant of poolside/Laguna-S-2.1.

Usage — TabbyAPI

Download the model with the Hugging Face CLI, then reference it by repo id:

hf download genevera/Laguna-S-2.1-DFlash-exl3 --local-dir models/genevera--Laguna-S-2.1-DFlash-exl3

yaml

# text_config.yml
model_name: <your-main-model-dir>
draft_model:
  draft_mode: model
  draft_model_name: genevera--Laguna-S-2.1-DFlash-exl3

Conversion details

Source: poolside/Laguna-S-2.1-DFlash (BF16, 2.2 GB)
Tool: exllamav3 convert_model (Python 3.10, fabric env)
Bits: 4.0 | Head bits: 6 | Shards: 8192 MB max (single 535 MB file)
Fits entirely on one GPU; TabbyAPI assigns devices automatically

Files

File	Size	Notes
`model.safetensors`	535 MB	EXL3 quantized weights
`config.json`	1.6 KB	Architecture config
`quantization_config.json`	61 KB	EXL3 quantization params
`tokenizer.json`	7.0 MB	Target tokenizer (from `poolside/Laguna-S-2.1`)