LiquidAI's LFM2.5 Runs a 128K Reasoning Model on MacBooks at 253 Tokens per Second
Liquid AI's LFM2.5-8B-A1B lands on Apple Silicon with 8-bit quantization, hitting 253 tok/s under 6 GB RAM for fully private on-device agents
Takeaways
- New release: Liquid AI released LFM2.5-8B-A1B-MLX-8bit, an 8-bit Apple Silicon build of their on-device reasoning model.
- Sparse MoE efficiency: 8.3B total parameters, only 1.5B active per token, fits in 9 GB RAM on Apple Silicon.
- Blazing speed: 253 tokens/s on M5 Max, 146 on Ryzen AI Max+ 395, ~30 tok/s on a phone, all under 6 GB.
- Massive benchmark gains: Non-hallucination rate jumped from 7.46% to 63.47%; MATH500 up from 74.8 to 88.76; Tau² Telecom up from 13.6 to 88.07 vs. predecessor.
- 128K context + reasoning: Context window quadrupled from 32K; model now produces explicit chain-of-thought before every answer.
- Limitations: Text-only (no vision/audio), not ideal for heavy code generation or knowledge-intensive QA without retrieval.
Running a reasoning model with a 128K context window on a MacBook at 253 tokens per second, under 6 GB of RAM, with no cloud dependency, was not a realistic expectation a year ago. LFM2.5-8B-A1B-MLX-8bit makes that the default experience. Liquid AI has released an 8-bit quantized, Apple Silicon-native build of their latest edge model, and the numbers are hard to ignore.
8B on the label, 1.5B on the clock
LFM2.5-8B-A1B is a sparse Mixture-of-Experts (MoE) model designed to run on consumer hardware while supporting tool-calling and long-context reasoning. The architecture totals 8.3B parameters but activates only 1.5B per token, enabling reasonable throughput and memory use on CPUs, laptops, and phones. That distinction is the whole story. MoE (Mixture-of-Experts) means the model is split into many specialized sub-networks called "experts," and only a small subset of them fire for any given token. You get the capacity of a large model at the compute cost of a much smaller one.
Active parameters let a model carry more latent capacity than a dense model of the same active budget, while making the deployed workload look closer to a smaller model. That is exactly what edge hardware wants. The model card describes LFM2.5-8B-A1B as a 24-layer model with 18 double-gated LIV convolution layers and 6 grouped-query attention layers. The config lists 32 experts, routes 4 experts per token, uses 32 attention heads, and uses 8 key-value heads for grouped-query attention.
What changed from LFM2
Compared to LFM2-8B-A1B, this new version expands the context window from 32,768 to 128,000 tokens, allowing the model to process longer documents and reason for longer. Its vocabulary size was also scaled up from 65,536 to 128,000 to tokenize non-Latin scripts more efficiently. Particularly strong compression gains appear in Hindi, Thai, Vietnamese, Indonesian, and Arabic.
The training pipeline also got a serious overhaul. LFM2.5 is a new family of hybrid models designed for on-device deployment, building on the LFM2 architecture with extended pre-training and reinforcement learning. Pretraining was scaled from 12T to 38T tokens. Unlike its predecessor, LFM2.5-8B-A1B is a reasoning-only model, producing an explicit chain of thought before its final answer. Liquid adopted this strategy because MoE models generally run in compute-bound settings, where a smaller number of active parameters makes each reasoning token cheap. This provides a significant quality boost without compromising speed.