Liquid AISpecialist modelOpen weights

LFM2.5-VL-3B-DSpark

PopularityRank 481 of 645

24out of 100-10 vs. last month

based on two data groups, mostly search, range 15 to 34

Summary

LFM2.5-VL-3B-DSpark is a speculative-decoding draft model. It proposes candidates for the target model LFM2.5-VL-3B, which verifies and corrects them when needed. The draft and target model form one inference system. DSpark accelerates local decoding for the target model, with open weights aimed at runtimes such as SGLang and MLX-VLM.

Liquid AI measures up to 2.66x decoding on an H100 and up to 3.13x on an Apple M5 Max under specified runtimes and settings. Image encoding and prefill do not become faster, so end-to-end gains can be smaller.

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 54 published benchmark results

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

M5 Max MLX Multi-turn decoding speedup (Show measurement, test conditions, and source)

Source value: 2.3
Score: 2.3
Metric: speedup over target-only inference
Unit: ratio
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-15
Methodology: Vendor-reported result. Target LiquidAI/LFM2.5-VL-3B plus draft LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16); MMSpec Multi-turn; M5 Max MLX-VLM FP16, batch 1, temperature 0, block 8, max 2,048 output tokens.
Context: Speedup is for the target-plus-draft system, not standalone DSpark quality or throughput.

Source Liquid AI DSpark Hugging Face model card

M5 Max MLX COCO decoding speedup (Show measurement, test conditions, and source)

Source value: 3.13
Score: 3.13
Metric: speedup over target-only inference
Unit: ratio
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-15
Methodology: Vendor-reported result. Target LiquidAI/LFM2.5-VL-3B plus draft LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16); MMSpec COCO; M5 Max MLX-VLM FP16, batch 1, temperature 0, block 8, max 2,048 output tokens.
Context: Speedup is for the target-plus-draft system, not standalone DSpark quality or throughput.

Source Liquid AI DSpark Hugging Face model card

M5 Max MLX TextVQA end-to-end speedup (Show measurement, test conditions, and source)

Source value: 1.56
Score: 1.56
Metric: speedup over target-only inference
Unit: ratio
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-15
Methodology: Vendor-reported result. Target LiquidAI/LFM2.5-VL-3B plus draft LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16); MMSpec TextVQA; M5 Max MLX-VLM FP16, batch 1, temperature 0, block 8, max 2,048 output tokens.
Context: Speedup is for the target-plus-draft system, not standalone DSpark quality or throughput.

Source Liquid AI DSpark Hugging Face model card

M5 Max MLX MMMU-Pro end-to-end speedup (Show measurement, test conditions, and source)

Source value: 2.62
Score: 2.62
Metric: speedup over target-only inference
Unit: ratio
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-15
Methodology: Vendor-reported result. Target LiquidAI/LFM2.5-VL-3B plus draft LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16); MMSpec MMMU-Pro; M5 Max MLX-VLM FP16, batch 1, temperature 0, block 8, max 2,048 output tokens.
Context: Speedup is for the target-plus-draft system, not standalone DSpark quality or throughput.

Source Liquid AI DSpark Hugging Face model card

H100 SGLang Multi-turn decoding speedup (Show measurement, test conditions, and source)

Source value: 2.04
Score: 2.04
Metric: speedup over target-only inference
Unit: ratio
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-15
Methodology: Vendor-reported result. Target LiquidAI/LFM2.5-VL-3B plus draft LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16); MMSpec Multi-turn; single H100 80GB SGLang BF16, batch 1, temperature 0, block 9.
Context: Speedup is for the target-plus-draft system, not standalone DSpark quality or throughput.

Source Liquid AI DSpark Hugging Face model card

H100 SGLang COCO decoding speedup (Show measurement, test conditions, and source)

Source value: 2.66
Score: 2.66
Metric: speedup over target-only inference
Unit: ratio
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-15
Methodology: Vendor-reported result. Target LiquidAI/LFM2.5-VL-3B plus draft LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16); MMSpec COCO; single H100 80GB SGLang BF16, batch 1, temperature 0, block 9.
Context: Speedup is for the target-plus-draft system, not standalone DSpark quality or throughput.

Source Liquid AI DSpark Hugging Face model card

H100 SGLang TextVQA end-to-end speedup (Show measurement, test conditions, and source)

Source value: 1.64
Score: 1.64
Metric: speedup over target-only inference
Unit: ratio
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-15
Methodology: Vendor-reported result. Target LiquidAI/LFM2.5-VL-3B plus draft LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16); MMSpec TextVQA; single H100 80GB SGLang BF16, batch 1, temperature 0, block 9.
Context: Speedup is for the target-plus-draft system, not standalone DSpark quality or throughput.

Source Liquid AI DSpark Hugging Face model card

H100 SGLang COCO end-to-end speedup (Show measurement, test conditions, and source)

Source value: 2.27
Score: 2.27
Metric: speedup over target-only inference
Unit: ratio
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-15
Methodology: Vendor-reported result. Target LiquidAI/LFM2.5-VL-3B plus draft LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16); MMSpec COCO; single H100 80GB SGLang BF16, batch 1, temperature 0, block 9.
Context: Speedup is for the target-plus-draft system, not standalone DSpark quality or throughput.

Source Liquid AI DSpark Hugging Face model card

H100 SGLang MMMU-Pro draft acceptance (Show measurement, test conditions, and source)

Source value: 4.11
Score: 4.11
Metric: mean draft tokens accepted per target verification pass
Unit: mean-count
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-18
Methodology: Vendor-reported result. Target model LiquidAI/LFM2.5-VL-3B; draft checkpoint LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16). Task row: MMMU-Pro. one H100 80 GB, SGLang, BF16, batch size 1, temperature 0, block size 9; source says both vision encoder and language backbone use 16-bit processing. This is the draft-acceptance table observation.
Context: Vendor-reported mean accepted draft tokens per target verification pass, not a standalone DSpark quality or throughput score and not itself an end-to-end speedup. The card says acceleration also depends on hardware and runtime. The source reports no per-cell sample count or confidence interval; table cells are correlated measurements, not independent experiments.

Source Liquid AI DSpark Hugging Face model card

M5 Max MLX-VLM MMMU-Pro draft acceptance (Show measurement, test conditions, and source)

Source value: 4.07
Score: 4.07
Metric: mean draft tokens accepted per target verification pass
Unit: mean-count
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-18
Methodology: Vendor-reported result. Target model LiquidAI/LFM2.5-VL-3B; draft checkpoint LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16). Task row: MMMU-Pro. Apple M5 Max, MLX-VLM, FP16 weights, batch size 1, temperature 0, block size 8, up to 2,048 output tokens; source says both vision encoder and language backbone use 16-bit processing. This is the draft-acceptance table observation.
Context: Vendor-reported mean accepted draft tokens per target verification pass, not a standalone DSpark quality or throughput score and not itself an end-to-end speedup. The card says acceleration also depends on hardware and runtime. The source reports no per-cell sample count or confidence interval; table cells are correlated measurements, not independent experiments.

Source Liquid AI DSpark Hugging Face model card

M3 Ultra llama.cpp MMMU-Pro draft acceptance (Show measurement, test conditions, and source)

Source value: 4.19
Score: 4.19
Metric: mean draft tokens accepted per target verification pass
Unit: mean-count
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-18
Methodology: Vendor-reported result. Target model LiquidAI/LFM2.5-VL-3B; draft checkpoint LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16). Task row: MMMU-Pro. Apple M3 Ultra, llama.cpp, FP16 weights, batch size 1, temperature 0, block size 8, up to 2,048 output tokens; source says both vision encoder and language backbone use 16-bit processing. This is the draft-acceptance table observation.
Context: Vendor-reported mean accepted draft tokens per target verification pass, not a standalone DSpark quality or throughput score and not itself an end-to-end speedup. The card says acceleration also depends on hardware and runtime. The source reports no per-cell sample count or confidence interval; table cells are correlated measurements, not independent experiments.

Source Liquid AI DSpark Hugging Face model card

H100 SGLang Multi-turn draft acceptance (Show measurement, test conditions, and source)

Source value: 3.46
Score: 3.46
Metric: mean draft tokens accepted per target verification pass
Unit: mean-count
Category: inference-performance
Direction: Higher is better
Source type: vendor-reported
Evaluator: Liquid AI DSpark Hugging Face model card
Status: active
Retrieved at: 2026-10-18
Methodology: Vendor-reported result. Target model LiquidAI/LFM2.5-VL-3B; draft checkpoint LiquidAI/LFM2.5-VL-3B-DSpark (279.5M BF16). Task row: Multi-turn. one H100 80 GB, SGLang, BF16, batch size 1, temperature 0, block size 9; source says both vision encoder and language backbone use 16-bit processing. This is the draft-acceptance table observation.
Context: Vendor-reported mean accepted draft tokens per target verification pass, not a standalone DSpark quality or throughput score and not itself an end-to-end speedup. The card says acceleration also depends on hardware and runtime. The source reports no per-cell sample count or confidence interval; table cells are correlated measurements, not independent experiments.

Source Liquid AI DSpark Hugging Face model card

Page 1 of 5

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

15 of 15 models

SAM 376

TeleOCR39

Agora-233

LFM2.5-VL-3B-DSpark24

ManiFlow-tb21

Molmo2-4B-tb21

pplx-embed-v2-context-9b-preview11

LFM2.5-VL-3B-DSpark

pplx-embed-v2-context-9b-…

FLUX 3 Action

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 9, 2026

LFM2.5-VL-3B-DSparkAgora-2 33LensVLM-9B 32AstaBrief-8B 30Hy-SimulMT-30B-A3B 27

Method changes in the timeline, a jump at these points is no market event:

1November 1, 2026: Value only with search or usage data, 30-day attention
2November 1, 2026: Method version 4 (common scale, shrinkage)
3November 9, 2026: New topics measurement V20 (daily quorum, short forms)

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

Similar models

More models of the same kind as LFM2.5-VL-3B-DSpark, chosen by model line, developer, size, price and release date.

All specifications

Architecture

Model class: Speculative-decoding draft model; Source: Model class

Input and output

Speed: Up to 2.66x faster decoding on H100 and 3.13x on Apple M5 Max, both vendor-measured; Source: Speed

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking: Unknown; no source
Watermarking technique: Not known; no source

Licence and access

Access: Local download, no standalone hosted API service; Source: Access
License: LFM 1.0; Source: License
Shutdown: None announced; no source

API

Model ID: LiquidAI/LFM2.5-VL-3B-DSpark; Source: Model ID

Further details

Target model: LiquidAI/LFM2.5-VL-3B; Source: Target model

Sources and data date

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

Type	Evidence and data date
Research status	Research date October 15, 2026. 2 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	standalone-quality-benchmark Note Official card identifies LFM2.5-VL-3B as target and DSpark as drafter; quality benchmarks are for target only, not a standalone draft score.
Not published	per-cell-sample-count-and-confidence-interval Note The model card gives no sample count or confidence interval for these table cells. The explicitly reported metric values are retained without inferred N or uncertainty.
Additional source	Liquid AI model card (retrieved September 26, 2026; October 15, 2026; October 18, 2026) · Liquid AI model card · Editorial description reviewed October 4, 2026
Additional source	Liquid AI announcement (retrieved September 26, 2026; October 15, 2026) · LFM2.5-VL-DSpark announcement · Editorial description reviewed October 4, 2026
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Ollama: cumulative downloads for the popularity index
Additional source	Replicate: cumulative runs for the popularity index
Additional source	Google Trending Now: limited feed selection for Germany and the US
Additional source	Creative Commons Attribution 4.0