Bielik-11B-v3.1-DFlash2

Bielik-11B-v3.1-DFlash2

Pairs with the Bielik-11B-v3.1-Instruct model as a draft model for speculative decoding in SGLang or vLLM. It is not a standalone language model.

Made by speakleash on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Bielik-11B-v3.1-DFlash2 is a small draft model for speculative decoding. It is meant to run next to Bielik-11B-v3.1-Instruct, the larger target model, inside an inference server that supports DFlash. On its own it does not work as a language model.

Who it suits: It suits teams that already serve the Bielik-11B-v3.1-Instruct model, mainly for Polish, and want to try speculative decoding with a matching draft.

Industry | Who uses it | How they use it | Example
Language technology | Polish NLP teams | Pair the draft with the Instruct model on their own server | Serving a Polish chat model through vLLM
Inference engineering | Platform engineers | Try DFlash speculative decoding in SGLang or vLLM | Comparing the two servers' launch setups
Research | Applied ML groups | Study draft models trained on the target's own data | Reading how the draft was trained on the same dataset

You can skip it if you do not run the Instruct model, because the draft does nothing alone.

Where to use it:

• SGLang, version 0.5.19 or newer
• vLLM, version 0.28.0 or newer
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Polish chat serving | Platform teams | A speculative decoding setup for the Instruct model
Server comparison | Inference engineers | One draft model tried in both SGLang and vLLM
Draft model study | Researchers | A draft trained on the same data as its target

How to set it up:

• Python, to run the Hugging Face download tool
• The huggingface_hub package
• SGLang 0.5.19 or newer, or vLLM 0.28.0 or newer
• The larger Instruct model that the draft pairs with

1. Install the Hugging Face hub package.
2. Download the draft model files to your computer.
3. Start SGLang or vLLM with the Instruct model as the main model and this one as the draft.

pip install -U huggingface_hub
hf download speakleash/Bielik-11B-v3.1-DFlash2

The card's SGLang launch looks like this.

export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1

python -m sglang.launch_server \
    --model-path speakleash/Bielik-11B-v3.1-Instruct \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path speakleash/Bielik-11B-v3.1-DFlash2 \
    --speculative-num-draft-tokens 16

The card's vLLM launch looks like this.

vllm serve speakleash/Bielik-11B-v3.1-Instruct \
  --speculative-config '{"method": "dflash", "model": "speakleash/Bielik-11B-v3.1-DFlash2", "num_speculative_tokens": 15}' \
  --attention-backend flash_attn \
  --max-num-batched-tokens 32768

How to use it: You start the server with the Instruct model as the main model and name this model as the draft. After that you use the server the way you already use the Instruct model.

Tool | What it does
python -m sglang.launch_server | Starts SGLang with the target model and the draft
--speculative-num-draft-tokens | Sets how many tokens the draft proposes, 16 in the card's example
vllm serve | Starts vLLM with the target model
--speculative-config | Names the dflash method, the draft model and 15 draft tokens

Example: You launch vLLM with the Instruct model and this draft in the speculative config, then chat with the model in Polish as before. The replies come from the Instruct model, and the draft only proposes tokens for it to check.

For newcomers and experts: A draft model is a small helper that guesses the next few words so a larger model can check them together. This one never writes answers alone. To start, download it and launch the vLLM example above.

If you know the field, this is a DFlash2 draft with 1.2B parameters, trained on the target's own dataset with a custom setup built on ALLaMo. The cost is that it ties you to one target model and to a server with DFlash support.

You send a request to the server
The draft proposes several tokens
The Instruct model checks them together
The accepted tokens form the reply

Good to know:

Title: Draft tokens in the card's two setups
SGLang setup: 16 draft tokens
vLLM setup: 15 draft tokens

• Not standalone: the card says it must be paired with its target model
• Languages: 32 European languages, tuned for Polish
• Size: about 2.5 GB in BF16, or about 617 MB at 4-bit as an estimate
• License: Apache 2.0
• Cost to run: none named, but you also need the Instruct model and a supported server
• Data it reaches: the prompts you send to your own server, and the files you download from Hugging Face
• Not for: standalone text generation, since it is only a draft

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 1.2B
• Weights: about 2.5 GB in BF16; about 617 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 44
• Likes: 3
• License: apache-2.0
• Languages: pl, en
• Quantized versions: 2
• Tags: transformers, safetensors, qwen3, dflash, speculative-decoding, block-diffusion, draft-model, efficiency, bielik, diffusion-language-model, text-generation, multilingual

Links

• Model: https://huggingface.co/speakleash/Bielik-11B-v3.1-DFlash2
• Files: https://huggingface.co/speakleash/Bielik-11B-v3.1-DFlash2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download speakleash/Bielik-11B-v3.1-DFlash2

Then follow the usage in its model card. Nothing in this kit runs by itself.