Liquid launches 2.6 billion-parameter agent model with 128 k token context window

Liquid announced LFM2.5-2.6B, a 2.69 billion-parameter model built with 30 layers, 22 short convolution blocks with dual gating and 8 grouped-query attention (GQA) blocks. The vocabulary was doubled to 128k tokens by extending the existing tokenizer rather than retraining, and a dedicated mid-training stage lengthened the context to 131072 tokens. The model was pre-trained on roughly 34 trillion tokens, covering 16 languages, and is text-only. Two checkpoints are released: a base version for fine-tuning and a version that underwent four-stage post-training for agent workloads.

Liquid benchmarked the model against Jamba-4-E2B-it (5.1 billion), Jamba-4-E4B-it (8 billion), Coven-3.5-4B (4.7 billion) and Coven-3.5-9B (9.7 billion). LFM2.5-2.6B leads all reported instruction-following tests and almost all tool-use tests, falling behind only Coven-3.5-9B on BFCLv4. In the LiveCodeBench v6 gap, it scores 59.41 versus 69.86 for Coven-3.5-9B. The company explicitly states the model is not recommended for coding agents or knowledge-intensive tasks.

Both checkpoints are available on Hugging Face under the lfm1.0 license, with open weights but no open-source code, in native, GGUF, MLX and ONNX formats, and immediate support in llama.cpp, vLLM, SGLang and LM Studio. On an M5 Max the model decodes 220 tokens per second using less than 2.5 GB of memory; on a phone the measured throughput is about 30 tokens per second. A single H100 GPU can serve roughly 1.3 billion tokens per day. Fine-tuning is supported via LoRA with TRL and Unsloth.

Liquid targets automotive, consumer electronics, industrial robotics, health, finance, e-commerce and security sectors, emphasizing regulated and offline environments where no prompt is sent to an external API. Practical use cases include local assistants, offline document classification on 128-k-token inputs, form and invoice extraction, robotics command analysis and continuously running background agents without per-token cost. Weights are released externally, performance figures are reported, and further development depends on downstream builders.