LFM2.5-VL-DSpark

Use it with LFM2.5-VL-3B to speed up vision-language model inference on supported NVIDIA GPU and Apple silicon runtimes.

Availability

Checkpoint available on Hugging Face; supported by SGLang, MLX-VLM, and llama.cpp via GGUF

Pricing model

Not established

Deployment

SGLang for NVIDIA GPUs, MLX-VLM for Apple silicon, or llama.cpp with GGUF

Privacy

Not established

Licence

Open-weight; the post says deploy without restrictions

• The draft model is available on Hugging Face in Safetensors and GGUF formats. ^{[1] [3]}

  Our vision DSpark draft model is available on Hugging Face in Safetensors and GGUF formats .

• LFM2.5-VL-3B-DSpark is a speculative-decoding draft model for LFM2.5-VL-3B. ^{[1] [2] [3]}

  **LFM2.5-VL-3B-DSpark** is an experimental speculative-decoding draft model that brings DSpark to

• Speculative decoding is exact under greedy decoding and preserves the target model’s output distribution under matched non-zero-temperature sampling settings. ^[2]

  Under matched sampling settings at non-zero temperatures, speculative decoding preserves the target model's output distribution.

• On one H100 with SGLang, decoding runs up to 2.66× faster. ^{[1] [2] [3]}

  In SGLang on a single H100, decoding runs up to 2.66× faster.

• On an Apple M5 Max with MLX-VLM, decoding runs up to 3.13× faster. ^{[1] [2] [3]}

  On Apple silicon, it reaches up to 3.13× with MLX-VLM on an M5 Max and up to 2.14× with llama.cpp on an M3 Ultra.

• The model is supported by SGLang for NVIDIA GPUs and MLX-VLM for Apple silicon; llama.cpp users can use its GGUF checkpoint. ^{[1] [2]}

  LFM2.5-VL-3B-DSpark is supported by SGLang for NVIDIA GPUs and MLX-VLM for Apple silicon.

• New models and tools focus on typed decisions and long-context tasks Weekly brief 2026-W39
• Claude Platform adds automatic evaluation for Managed Agents tool calls Daily brief 2026-09-24