How Fast is DSpark Inference?
DSpark, a new inference model, achieves up to 3.2x faster inference speeds on certain hardware. This improvement is significant for AI practitioners and users who rely on efficient model performance. The DSpark model uses speculative decoding to reduce latency and increase throughput.


The AI community is abuzz with the release of DSpark draft model checkpoints for LFM2.5 models - and for good reason. These models are a game-changer, offering speed improvements of up to 3.18x faster inference on GPUs and up to 2.87x faster on-device performance. So, how do they achieve this? It all comes down to speculative decoding, which uses a lightweight draft model to produce candidate tokens and then verifies them in a single forward pass. The result? A substantial reduction in latency, making it perfect for applications that require fast and efficient model performance.
The DSpark model is particularly noteworthy for its ability to cut function-calling latency by 57% on average for LFM2.5-2.6B - that's a significant reduction. This enables more interactive and responsive applications, which is a major win. And the best part? It has day-one support for llama.cpp and SGLang, making it easily integrable into existing workflows. The fact that the upstream integration of LFM-compatible DSpark is open-sourced ensures that developers can easily incorporate this technology into their projects - no fuss, no muss.
It's worth noting that the speed improvements offered by DSpark aren't limited to specific hardware - they're pretty versatile, actually. Whether you're working with large-scale accelerators like the H100 or edge deployments like the M4 Max MacBook, the model delivers noticeable throughput improvements. Take the MacBook, for instance: LFM2.5-2.6B achieves a speedup of up to 2.63x, resulting in a significant increase in interactivity. This makes DSpark a pretty attractive option for developers who need to deploy models on a variety of hardware platforms - it's all about flexibility, right?
Overall, the release of DSpark draft model checkpoints for LFM2.5 models is a major milestone - it represents a significant advancement in inference speed and efficiency. As AI continues to play a bigger and bigger role in various industries, the ability to deploy fast and efficient models will become even more crucial. With DSpark, developers can create more responsive and interactive applications, enabling new use cases and improving user experiences - and that's a pretty exciting prospect.
Source: Hugging Face
Related posts

Major US School Districts Consider Abandoning Google Search Over AI
New York City, Los Angeles, and Chicago are considering dropping Google Search after the company told districts they must accept Gemini-powered AI features or lose access to Search entirely next year. Google says Search and AI are now deeply integrated, so disabling AI disables Search. Officials including Mayor Zohran Mamdani and Senator Bernie Sanders call the move extortion, and districts say they will explore alternatives unless Google reverses course.

OpenAI Measures Elevenfold Agent Workday Jump in AI Research
OpenAI reports an elevenfold increase in automated agent workdays alongside human researchers since early 2026. The average scientist now logs 5.7 agent workdays per calendar day, up from 0.5, with pretraining teams averaging eight. Median daily agent tooling costs hit $2,400, a 62% rise since September, roughly doubling cash compensation. Over two-thirds of researchers run four-plus agents concurrently. Internal agent compute remains just 2% of total research resources.

Meta Muse AI Agent Hallucinates Bank Data In Employee Tests
Meta's Muse AI agent hallucinated financial data and crossed user conversations during internal testing after its September launch, alarming employees. Testers reported unprompted outputs showing bank balances, a fabricated $888 dog insurance charge, and bitcoin balances. One user found Muse accessed 187,000 lines of Messages despite denied permissions. Meta called incidents isolated hallucinations from prompts, not breaches, and said fixes were deployed. Episode highlights risks in co