Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official4
liquid.ai2026-08-21about the LFM2.5 familyHN 18 pts · 0 comments1,933 words
Liquid AI releases DSpark draft models for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B, integrating speculative decoding via llama.cpp and SGLang. The draft models use 5 layers with a block size of 9 and total approximately ~300M parameters, achieving up to 3.18x GPU speedup and up to 2.87x on-device gains for the 2.6B variant. While the 1.2B and 2.6B models see consistent improvements, the 8B-A1B variant only achieves an 18% improvement on edge devices due to current MoE implementation constraints in llama.cpp's Metal backend.written, reviewed, cited
- The 8B-A1B draft model adds 65.5M parameters for its Markov head compared to 33.6M in the 1.2B and 2.6B variants.
- For the 1.2B-Instruct draft, validation loss consistently tracks acceptance rate improvements across all training epochs.
- The 8B-A1B draft model achieves only an 18% improvement on edge devices because verifying k tokens activates more experts and weight traffic.
huggingface.co2026-08-06authorityabout the LFM2.5 familyHN 1 pts · 0 comments1,103 words
Liquid AI releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, two new models designed for fast long-context inference on CPU. These encoders support an 8,192-token context and are about 3.7× faster than ModernBERT-base at long context lengths. The models are trained with a masked language objective that masks 30% of tokens during training to achieve strong performance on GLUE, SuperGLUE, and multilingual tasks.written, reviewed, cited
- The 350M variant is the choice when accuracy matters most.
- The 230M variant is the choice for tighter hardware or higher throughput.
- A base encoder provides general-purpose representations rather than direct task outputs.
docs.liquid.ai2026-08-04HN 3 pts · 0 comments161 words
LFM2.5-2.6B is Liquid AI's 2.6B dense model built for agentic workloads with a 131K context window and native tool calling. It runs on edge devices and is trained to work reliably inside agent harnesses like Hermes Agent, OpenClaw, and Pi. The chat template opens every assistant turn with a <think> tag so responses contain an explicit chain of thought before the final answer. Users can run it via Transformers, llama.cpp, vLLM, or SGLang.written, reviewed, cited
liquid.ai2026-08-04HN 3 pts · 0 comments1,566 words
Liquid AI releases LFM2.5-2.6B, a 2.6B parameter agentic model trained on ~34T tokens and available on Hugging Face. It features a 128K vocabulary and context-extension phase, decoding 220 tokens/s on an M5 Max and 113 tokens/s on a Ryzen AI Max+ 395 while staying under 2.5 GB. The model reaches 15K output tokens per second on an H100 GPU.written, reviewed, cited
- The model doubles its vocabulary to 128K by extending the existing tokenizer in place rather than retraining the model from scratch.
- Training uses a dedicated 128K context-extension phase so the model can handle the long inputs that agentic workloads require.
- The training pipeline separates model optimization, inference, and environment execution into distinct components using FSDP, SGLang, and verl.
