LFM2.5-DSpark Release Notes / What's New

Liquid AI has released DSpark draft model checkpoints for three models in the LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. This update introduces a speculative decoding path that significantly increases decoding speed and reduces function-calling latency while maintaining identical output quality to the baseline models.

Technical Architecture of DSpark

DSpark optimizes the decode phase of LLM inference, which is typically memory-bound due to the cost of streaming weights from DRAM to SRAM. It utilizes a lightweight draft model to propose candidate tokens that the target model then verifies in a single forward pass.

DSpark integrates three primary components to achieve this:

  • Parallel Backbone: A DFlash-style backbone conditioned on the target model’s context features, producing hidden states for all draft tokens in one forward pass.
  • Sequential Head: A lightweight Markov chain-based head that adds inter-token dependency to increase the acceptance rate of tokens at later positions.
  • Confidence-Scheduled Verifier: A mechanism that predicts the survival probability of each token and prunes low-confidence suffixes to avoid verification costs that exceed potential savings.

Draft Model Training and Specifications

Liquid AI trained the draft models using a diverse data mix including SFT, chat, code, and function-calling data. The models are simplified attention-only architectures consisting of 5 layers and a block of 9. Training lasted 15 epochs, with the final checkpoints selected based on the highest acceptance rate rather than the lowest loss.

Each draft model is approximately 300M parameters. The specific parameter breakdowns are as follows:

Component LFM2.5-1.2B-Instruct LFM2.5-8B-A1B LFM2.5-2.6B
Decoder stack (5 layers) 241.2M 241.2M 241.2M
Hidden-state projection 21.0M 21.0M 21.0M
Markov head 33.6M 65.5M 65.5M
Norms + confidence head 27.5k 27.5k 27.5k
Total 295.7M 327.7M 327.7M

Performance Benchmarks

Because DSpark uses greedy decoding where a draft token is only accepted if it matches the target model’s distribution, the output is identical to baseline greedy decoding. Consequently, benchmark accuracy (pass@1 or exact match) remains unchanged.

GPU and On-Device Throughput

Testing was conducted using SGLang on a single H100 80 GB (BF16) and llama.cpp with Metal on an M4 Max MacBook Pro (FP16 GGUF). All configurations used a batch size of 1, a temperature of 0, and a DSpark block size of 9.

LFM2.5-2.6B Performance

Dataset Acceptance (of 10) Speedup on H100 Speedup on M4 Max
MATH500 5.42 3.06x 2.25x
HumanEval 4.54 2.56x 2.63x
MBPP 4.71 2.64x 2.11x
GSM8K 4.32 2.22x 2.36x
MT-Bench 5.07 2.87x 1.99x
Mean 4.81 2.67x 2.27x

LFM2.5-1.2B-Instruct Performance

Dataset Acceptance (of 10) Speedup on H100 Speedup on M4 Max
MATH500 6.02 2.56x 2.62x
HumanEval 5.31 2.26x 2.87x
MBPP 5.52 2.37x 2.74x
GSM8K 4.34 1.67x 2.73x
MT-Bench 3.90 1.66x 1.72x
Mean 5.02 2.10x 2.54x

LFM2.5-8B-A1B Performance

Dataset Acceptance (of 10) Speedup on H100 Speedup on M4 Max
MATH500 8.27 3.18x 1.21x
HumanEval 7.02 2.58x 1.12x
MBPP 6.93 2.64x 1.09x
GSM8K 4.02 1.29x 1.44x
MT-Bench 8.52 3.02x 1.04x
Mean 6.95 2.54x 1.18x

For the LFM2.5-8B-A1B model, on-device speedup is limited to an average of 18% due to current MoE implementation constraints in llama.cpp’s Metal backend and the increased weight traffic required to verify multiple tokens.

Agentic Inference Latency

For the LFM2.5-2.6B model, DSpark reduces function-calling latency by an average of 57% across various multi-tool scenarios.

Implementation and Deployment

LFM2.5-DSpark is supported on day one via the following frameworks:

  • SGLang: Requires a build with DSpark support for LFM2 targets (PR #31041).
  • llama.cpp: Requires a specific build (PR #27383).

Draft model checkpoints are available on Hugging Face in both Safetensors and GGUF formats for the 1.2B-Instruct, 2.6B, and 8B-A1B variants.

Sources

Related

  • Dispatch
    vLLM speculative decoding on AMD GPUs: performance and methods

    vLLM adds speculative decoding to AMD Instinct MI300X/MI355X GPUs, letting a lightweight draft model propose multiple tokens that the target model verifies in a single pass, which can double or more the output-token throughput depending on the draft method, model family, and proposal length.

  • Dispatch
    LFM2.5-2.6B release notes / what's new

    Liquid AI has released LFM2.5-2.6B, a small, high-performance model designed for on-device agents with best-in-class tool use and instruction following capabilities.

  • Dispatch
    LFM2.5-Encoders Release

    Liquid AI has released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, general-purpose encoder models that provide high-quality long-context inference (up to 8,192 tokens) with significantly faster CPU performance than ModernBERT-base.

  • Dispatch
    LFM2.5-VL-3B release notes / what's new

    Liquid AI has released LFM2.5-VL-3B, a 3.1B parameter vision-language model optimized for edge devices, featuring improved screen understanding, grounding, and tool calling.

  • Dispatch
    LFM2.5 Q4_0 release notes / what's new

    Liquid AI has released Q4_0 GGUF checkpoints for LFM2.5 models using Quantization-Aware Distillation (QAD) to recover 97% of the accuracy lost to quantization while maintaining 4-bit memory and speed.