Build efficient general-purpose AI at every scale.
Cambridge, MA
- Tweets 1,016
- Following 59
- Followers 36,526
- Likes 2,495
Today we release LFM2.5-2.6B, an agentic model that runs entirely on-device. It plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots. Data never leaves the device, and the marginal cost of each run is essentially zero.
> Pre-trained on ~34T tokens
> LFM2.5 flagship hybrid architecture
> Context length: 128K
> Vocab size: 128K
> balanced intelligence per watt
> customizable on a single GPU for any specialized task
> LFM2 open-weight license
Comparable or better scores compared to models up to nearly 4x its size:
> ToolSandbox 77.83, ahead of Qwen3.5-9B at 76.44
> Multi-IF 80.07, ahead of Gemma-4-E4B-it at 77.35
> IFStruct 85.49, ahead of Qwen3.5-9B at 78.50
🧵
204
530
4,267
791,505
Liquid AI retweeted
DSpark for the already fast LFM vision models 🚀
Very useful for local image processing workflows
our VLMs just got a speedup!
DSpark speculative decoding for LFM2.5-VL-3B is here!
huggingface.co/LiquidAI/LFM2…
2
2
25
3,397
Liquid AI retweeted
Run LFM2.5-VL-3B on 16GB Mac with DSpark ⚡️
@liquidai's new DSpark drafter delivers up to 3.13x faster decoding on M5 chips with MLX-VLM in its vision benchmarks while keeping identical outputs under greedy decoding
Run AI models locally -> atomic.chat
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
6
4
53
5,912
magic ✨
Next agentic experience on Day 3 of #snapdragonsummit coming from @liquidai. Unbelievable how context and memory shared across devices can empower agents to be proactive and deeply personal. Magic
1
3
30
2,334
What happens when you bring DSpark to vision-language inference?
I ran LFM2.5-VL-3B with DSpark through llama.cpp on an M3 Pro, using a BF16 target.
Decode throughput went from 23.0 → 48.7 tok/s, giving a 2.12× speedup , very close to Liquid AI reported results.
Nice to see the gains hold up on a local Apple Silicon setup.
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
5
2
24
2,278
Liquid AI retweeted
Our latest VLM now decodes up to 3x faster:
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
2
22
2,118
Liquid AI retweeted
Check out our release of a draft model for LFM2.5-VL-3B:
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
1
2
18
1,701
Liquid AI retweeted
Congratulations to the @liquidai team for releasing DSpark draft model for LFM2.5-VL-3B! 🚀
Excited to have partnered with them for day-0 support on MLX-VLM and @Nativ_AI.
This draft model achieves decoding throughput improvements of up to 3.13× on edge devices, with end-to-end throughput gains of up to 2.27× and 2.62×.
Special thanks to @tugot17, he cooked!
HF weights 👇🏽
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
5
4
60
4,810
Liquid AI retweeted
fast in every scale💨💨
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
2
16
1,648
Our partners at @LiquidAI brought Liquid Context to Snapdragon: personal context that stays on the device and is handed to agents, local or cloud.
Same layer we build for the Mac: Mnemos keeps memory on the Mac, models plug in below.
My bet: people pick a device for its context.
2
1
11
803
Liquid AI retweeted
after text now we are making vision faster as well 📷 🫡
probably the first speculator for small VLMs if I'm not mistaken
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
4
5
37
2,662
Liquid AI retweeted
M1 16G MacBook Pro Saver! 😑
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
4
18
2,033
Liquid AI retweeted
our VLMs just got a speedup!
DSpark speculative decoding for LFM2.5-VL-3B is here!
huggingface.co/LiquidAI/LFM2…
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
8
9
137
9,214
Liquid AI retweeted
Vision-language inference just got faster ⚡️
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
1
2
21
3,190
Liquid AI retweeted
a lightweight drafter for vision language models 🌊✨
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
2
3
34
2,300
Today, we release an experimental DSpark draft model for LFM2.5-VL-3B, bringing speculative decoding to our vision-language models.
A lightweight drafter proposes multiple tokens ahead, and the target model verifies them together in a single pass. This accelerates generation without changing output quality.
Across six vision-language task categories at batch size 1 and temperature 0:
> MLX on M5 Max: up to 3.13x faster decoding and 2.62x end to end
> llama.cpp on M3 Ultra: up to 2.14x faster decoding and 1.77x end to end
> SGLang on H100: up to 2.66x faster decoding and 2.27x end to end
All evaluations were collected using Pipette, the benchmarking infrastructure behind Liquid AI's public device-performance data.
🧵
23
63
475
50,832
The resulting drafter adds 279.5M parameters, increasing the deployed model's parameter count by 8.9%. This total excludes the embedding and LM head, which are tied to the target model rather than carried by the drafter.
After data-mixture and architecture ablations we selected:
> A 4-layer, attention-only drafter
> Block size 9 during training
> Block size 8 or 9 at inference, depending on hardware
Every drafter experiment and ablation was trained on @AMD hardware using Liquid's training framework.
2
3
25
1,292
Vision workloads also show where speculative decoding stops helping.
A VLM must first process pixels through a vision encoder and prefill the language backbone with hundreds of visual tokens. DSpark accelerates decoding, not vision encoding or prefill, so end-to-end gains depend on how much time each task spends generating tokens.
The LFM2.5-VL-3B DSpark checkpoints are available now on Hugging Face, with support for llama.cpp, MLX-VLM, and SGLang.
> Blog: liquid.ai/blog/lfm2-5-vl-dsp…
> LiquidAI/LFM2.5-VL-3B-DSpark: huggingface.co/LiquidAI/LFM2…
> LiquidAI/LFM2.5-VL-3B-DSpark-GGUF: huggingface.co/LiquidAI/LFM2…
> llama.cpp: github.com/ggml-org/llama.cp…
> MLX-VLM: github.com/Blaizzy/mlx-vlm/p…
> SGLang: github.com/sgl-project/sglan…
21
962
Liquid AI retweeted
Check out our latest Substack issue liquidai.substack.com/p/cont…
Recap on:
> Liquid Context, our on-device context layer for edge agents on Snapdragon, in partnership with @Qualcomm
> LFMs that match frontier models at predicting aging, in collaboration with @InSilicoMeds
> Great interview with our CEO on @SAIRfoundation's podcast
1
7
595
Liquid AI retweeted
#SnapdragonSummit day 2 is live, tune in now: nitter.net/i/broadcasts/1qKVmyrZE…
1,038
31
121
149,937