What Liquid DSpark’s “open” release actually permits
The weights are downloadable and the acceleration technique is inspectable. The license still places a commercial-use condition on larger organizations.
Primary sources ↓ · Limitations ↓
SHORT ANSWER
Start here.
Liquid released a 279.5-million-parameter experimental draft model designed to work with LFM2.5-VL-3B during speculative decoding. Its weights and integrations are available, but its LFM Open License v1.0 requires a separate commercial license for an organization with at least $10 million in annual revenue. “Downloadable weights” therefore does not mean unrestricted commercial use or fully open source.
- Bright publication
- Source review
- Reading question
- Which artifact is available, what does it accelerate, and who may use it commercially?
The drafter and the base model are a pair
DSpark does not replace the 3-billion-parameter vision-language model. It proposes candidate output tokens; the base model verifies them and preserves the final model distribution. The 279.5-million-parameter drafter is labeled experimental and must be paired with LFM2.5-VL-3B.
Liquid reports up to 3.13× faster decoding on an Apple M5 Max and 2.66× on an NVIDIA H100. Those are company benchmarks for the token-generation stage. They do not establish the same improvement for image encoding, prompt processing, every sampling mode, or an entire user request.
Sources: Introducing LFM2.5-VL-3B-DSpark · LFM2.5-VL-3B-DSpark model card
Code, weights and data answer different questions
The public model repository provides downloadable weights and documentation, while the release names llama.cpp, MLX-VLM and SGLang integrations. Each runtime has its own code and terms. None of that, by itself, publishes the training data or a complete reproducible training recipe.
The useful inspection is artifact by artifact: the drafter weights, base-model weights, runtime code, license, model card, evaluation method and training-data account. One open component does not transfer its permissions to the others.
Sources: LFM2.5-VL-3B-DSpark model card · LFM2.5-VL-3B-DSpark LICENSE
Read the license before the launch language
The repository’s own LICENSE is the operative document Bright found for the DSpark weights. It allows many uses but defines a commercial-use restriction: organizations with annual revenue of $10 million or more need a separate license from Liquid AI. That condition matters even when release copy uses broader language about access or use.
Bright therefore describes DSpark as an open-weight release under a source-available license with a commercial threshold. It does not call the model open source, certify the license, or infer rights for data and third-party runtime code.
Sources: LFM2.5-VL-3B-DSpark LICENSE · Introducing LFM2.5-VL-3B-DSpark
What this guide does not establish
- The speed figures are Liquid AI’s decode-only benchmarks and were not independently reproduced by Bright.
- MLX support in the cited release is limited to greedy sampling, and performance depends on hardware, runtime, prompt and generation settings.
- Bright is reporting the repository terms checked on September 25, 2026, not providing legal advice or certifying license compatibility.
- Public weights and inference integrations do not establish that training data, training code or a complete training recipe are available.
Primary sources
LFM2.5-VL-3B-DSpark LICENSE ↗
The DSpark repository’s copy of LFM Open License v1.0, including its commercial-use threshold.
LFM2.5-VL-3B-DSpark model card ↗
Release status, paired base model, parameter count, integrations and experimental limitations.
Introducing LFM2.5-VL-3B-DSpark ↗
Liquid AI’s release explanation and company-reported decoding benchmarks.
Source-reviewed 2026-09-25. Source review means Bright compared this explanation with the linked records; it is not independent replication or expert peer review by Bright.