Liquid AI releases LFM2-24B-A2B: 24B parameter MoE open source model, 32GB memory can run locally

Liquid AI recently released its largest open source model, LFM2-24B-A2B. This model using sparse hybrid expert architecture requires only 32GB of memory to run and achieves decoding speeds of 20-30 tok/s on consumer laptops and desktop devices, providing a new option for AI localization deployment.

Core conclusion

On May 2, 2026, Liquid AI’s LFM2-24B-A2B model appeared on the homepage of Hacker News and sparked heated discussions. This is a sparse hybrid expert architecture model with 24B total parameters (2.3B active parameters/Token). It adopts a convolution-first design concept and can be run in a 32GB memory environment. The CPU inference speed can reach 20-30 tok/s.

Key Points

Event Time: May 2, 2026 (hotly discussed on HN homepage), the model has been released earlier
Target group: AI application developers, local deployment enthusiasts, edge computing practitioners
Core Value: Open Source Weight + Runnable on Consumer Grade Hardware + Low Inference Latency

Background: The road to expansion of the LFM2 architecture

Liquid AI's LFM2 series adopts a hybrid architecture that is different from traditional Transformers - combining efficient gated short convolution blocks with a small number of group query attention (GQA) layers, optimized through hardware-in-the-loop architecture search. The LFM2-24B-A2B is the largest model in the series, scaling from a dense model of 350M all the way to 24B MoE, covering a parameter range of nearly two orders of magnitude.

This expansion follows a clear recipe: Deeper Layers + More Experts. The model increases from 24 to 40 layers and the number of experts per set of MoE blocks doubles from 32 to 64, while maintaining the top-4 routing policy. The key point is that the active parameters only grew from 1.5B to 2.3B (~1.5x), while the total parameters grew 3x (8.3B → 24B), and the inference latency and energy consumption always tracked the smaller active paths.

Key Impact (by Dimension)

Dimensions	Changes	Meaning for developers	Recommended actions
Deployment threshold	32GB memory can run, supports CPU/GPU/NPU	Say goodbye to high-end GPU dependence, can be deployed on MacBook and mini hosts	Test the running performance of LFM2-24B-A2B on local devices
Inference cost	Only 2.3B active parameters, low power consumption	Edge device inference cost close to 2B dense model	Evaluate alternatives to larger active parameter models such as Qwen3-30B-A3B
Open source ecosystem	Apache 2.0 style open source weights, GGUF/llama.cpp ready to use	No API Key required, fully localized	Integrated into existing llama.cpp/vLLM inference pipeline
Community feedback	HN community measured CPU inference 20-30 tok/s	A practical choice in a GPU-free scenario	Compare the cost performance of similar models such as Gemma4 / Qwen3.5

Architecture breakthrough: convolution-first MoE design

The most noteworthy innovation of LFM2-24B-A2B is its 1:3 attention to convolution ratio - only 10 of the 40 layers use GQA attention, and the remaining 30 layers are efficient convolution blocks. This is completely different from the current mainstream LLM design that almost exclusively uses attention layers.

This architectural choice brings two core advantages:

Low memory overhead: The KV cache requirement of the convolutional layer is much lower than that of the attention layer, suitable for consumer-grade devices with 32GB memory
High throughput service: When using vLLM continuous batch processing on an H100 SXM5, it reaches about 26.8K tokens/s under 1024 concurrent requests, surpassing the MoE model of the same scale

The model fully supports llama.cpp (GGUF multi-quantization options: Q4_0/Q4_K_M/Q5_K_M/Q6_K/Q8_0/F16), vLLM and SGLang, providing a one-stop inference solution from development to production.

Benchmark performance

On standard benchmarks such as GPQA Diamond, MMLU-Pro, IFEval, IFBench, GSM8K, and MATH-500, quality improves logarithmically linearly with total parameters, demonstrating that the LFM2 hybrid architecture follows predictable scaling behavior. There is no ceiling effect at smaller model sizes over the nearly 100x parameter span from 350M to 24B.

It is worth noting that Liquid AI chose to release the instruct version instead of the reasoning version, on the grounds that training is faster and the instruct model is more popular with the community. Pre-training is still in progress (17T tokens trained), LFM2.5-24B-A2B with reinforcement learning post-training will be launched once completed.

LFM2 benchmark performance chart

Community actual test feedback

Feedback from the Hacker News community on the LFM2-24B-A2B focused on its usefulness:

CPU inference is possible: DDR4 memory + CPU can run at a speed of 20-30 tok/s
GPU User Recommendation: If you have a GPU with 4GB+ VRAM, the community recommends dense models such as Gemma4 or Qwen3.5/3.6 as an alternative
Benchmark Call: The community wants to see more independent third-party quality comparisons, not just speed metrics

Tool entry

The following tool names naturally appear in the text, and the platform side will match the maintained tools library: Liquid AI, llama.cpp, vLLM, Hugging Face, DeepSeek, Gemini, Gemma, Qwen

Adaptation suggestions

Test the Q4_K_M quantized version of LFM2-24B-A2B in llama.cpp to evaluate local inference scenarios
For links in the AI Agent workflow that require local reasoning (such as content extraction and classification), this model can be considered to replace API calls.
Pay attention to subsequent versions of LFM2.5, training after reinforcement learning may lead to quality improvements

Reference sources

Internal link guidance

Want to learn local AI model deployment? Watch: DeepSeek V4 vs Claude Code: 90% Cheaper, Same Quality
Real AI tool application cases: Claude Code 48 hours to start a business: one person + US$29 monthly fee, monthly income in 3 months $9,000