On August 20th, Liquid AI released a set of LFM2.5-DSpark draft models on Hugging Face, adding speculative decoding to three of its own models: LFM2.5-1.2B-Instruct, 2.6B, and 8B-A1B. The company says the speed-up reaches up to 3.18x, with output identical to the original models.
The idea behind speculative decoding isn't complicated: a small model guesses the next few tokens first, so the large model doesn't have to compute them one at a time — instead it verifies the whole batch of candidates in a single forward pass. Correct guesses get used immediately; wrong ones get recomputed. What's saved is all the time otherwise spent shuttling weights in and out.
655MB Buys You Two to Three Times the Speed
The draft models Liquid AI paired with its main models this time are tiny. The 1.2B version uses a 295.7-million-parameter draft model, while the 2.6B and 8B-A1B share a 327.7-million-parameter one. The architecture is five layers of full attention, a 2048 hidden dimension, a 6144 intermediate layer, 32 query heads sharing 8 KV groups, and 9 candidate tokens proposed at a time. At BF16 precision, it uses 655MB of memory.
Rough math on a 16GB Mac: 655MB is about 4% of total memory, in exchange for local chat responses more than twice as fast. That trade-off is a good deal for on-device scenarios — what's scarce on-device was never peak compute, it's the few seconds of blank waiting.
The Numbers Depend on the Hardware
The official benchmarks cover two environments, and the results differ substantially.
| Target model | H100 avg | H100 peak | M4 Max avg | M4 Max peak |
|---|---|---|---|---|
| 1.2B-Instruct | 2.10x | 2.56x | 2.54x | 2.87x |
| 2.6B | 2.67x | 3.06x | 2.27x | 2.63x |
| 8B-A1B | 2.54x | 3.18x | 1.18x | 1.44x |
The headline 3.18x figure comes from 8B-A1B's best result on H100. The same model drops to just 1.18x on an M4 Max laptop. Liquid AI spelled out the reason itself: llama.cpp's Metal backend currently doesn't implement mixture-of-experts architectures well, making on-device MoE the weak point of this approach.
Acceptance-rate data was also published. For 8B-A1B, an average of 8.27 out of every 10 candidate tokens are accepted on MATH500, versus just 4.02 on GSM8K. Well-structured, predictable text like mathematical proofs is easier for the draft model to guess correctly; switch to loosely worded elementary word problems and the hit rate drops by half.
Another data point hits closer to practical use: in multi-tool function-calling scenarios, latency for the 2.6B version dropped by an average of 57%. Every step an agent takes waits for the model to emit a structured call, and these round trips are the most vulnerable to compounding latency.
Where the Method Came From
The name DSpark wasn't coined by Liquid AI. It comes from a paper DeepSeek open-sourced jointly with Peking University in late June, which pairs a parallel draft backbone with a lightweight Markov head, plus a mechanism that schedules verification length against real-time GPU load. At the time, the reported numbers were a 60% to 85% improvement in single-user generation speed, with per-GPU throughput reaching up to 6.6x under strict latency targets.
Two months later, the method showed up in a release from a US on-device model company. SGLang enables it with a --speculative-algorithm DSPARK flag, and llama.cpp supported FP16 GGUF on the Metal backend from day one. An upstream framework writing the algorithm's name directly into a command-line switch value is itself a sign the method has already become a de facto standard.
Chinese labs have open-sourced plenty of inference-side engineering work over the past couple of years, and the discussion has mostly centered on whether it saves on GPUs. DSpark's spillover this time delivers a different kind of payoff: the method has made its way into someone else's default path, and every deployment built on that framework from here on is running your idea.
The License Draws a Line
Weights are released in both Safetensors and GGUF formats, under the LFM Open License 1.0: entities with annual revenue under $10 million can use it commercially for free, while those above that line need to negotiate a separate commercial agreement. The only deployment option is self-hosting — Liquid AI doesn't offer a hosted API.
That line is drawn squarely between startups and mid-to-large enterprises, and the intent is fairly clear: trade a free tier for developers' default choice, and reserve the revenue for whoever can afford to pay.
Sources: Hugging Face official blog, Liquid AI model card, CocoLoop, MarkTechPost; speed-up multiples, parameter counts, and acceptance-rate data follow the official benchmark tables, license terms follow the model page.