LFM2 24B A2B

A Hugging Face community discussion thread titled "First Thoughts on the LFM2 24B A2B Release," opened by user tanyiades on February 28, captures early community impressions of the 24B parameter model with the new 2B-active-parameter MoE architecture. The poster describes real-world usage as feeling fast, smooth, and stable, with particularly strong responsiveness in low-latency interactive scenarios, and notes that the foundation appears solid while leaving significant headroom for further fine-tuning and post-training. The thread frames the model as a step forward relative to expectations closer to a 16B setup and credits the Liquid AI team's direction and progress, but it does not include benchmarks, ablations, or technical measurements. The page exposes community comment controls (sign in to comment, drag-to-upload media) and shows 6 reactions, indicating lightweight engagement rather than quantitative evaluation. No first-party release notes or changelog are included on the discussion page itself.

Artificial Analysis profiles LFM2 24B A2B as an open-weights, non-reasoning Liquid AI model released in February 2026, scoring 6 on the Artificial Analysis Intelligence Index, below the median (6) for open-weight non-reasoning models of similar size. Technical specifications confirm 23.8B total parameters, 2.3B active parameters per token, a 33K-token context window (about 49 A4 pages of size-12 Arial), text input and output, no reasoning mode on the evaluated page, and an lfm 1.0 license with weights hosted on Hugging Face. The pricing snapshot lists $0.00 per 1M input and output tokens, well below the cohort medians of $0.05 and $0.15. The comparison summary characterizes the model as below average on intelligence but well-priced within its peer set, and notes that a reasoning variant may also exist separately. Output speed, verbosity, and cost-per-Intelligence-task figures are marked unknown, limiting cross-axis comparison on the page. The model is grouped with the small class (4B–40B parameters) of open-weights peers for benchmark normalization.

Benchable documents LFM2-24B-A2B as a 24B-parameter Mixture-of-Experts model with 2B active parameters per token, released February 25, 2026, and the largest entry in the LFM2 family of hybrid architectures aimed at efficient on-device deployment. It supports a 32K context window, lists Hugging Face integration, and records pricing at $0.03 per 1M prompt tokens and $0.12 per 1M completion tokens via a third-party endpoint. The page's benchmark summary places the model in the 97th percentile for speed, with a 98% reliability success rate, and rates it 90th percentile on pricing competitiveness. Specific category scores on Benchable show General Knowledge at 97.5% and Ethics at 98.0%, while Instruction Following (38.0%), Reasoning (50.0%), and Coding (71.0%) trail, and hallucination-handling accuracy sits at 88.0%, indicating room for improvement on complex prompts. The model tops Email Classification and ranks in the top three for General Knowledge among models of comparable speed. Together is listed as the only available hosted endpoint for the 2026-02-24 snapshot, with the documented pricing carried through unchanged.

Liquid AI's official Models index page lists two dated entries that explicitly name the LFM2-24B-A2B variant: "LFM2-24B-A2B: Scaling Up the LFM2 Architecture" on 01.20.2026, and "No Cloud, No Waiting: Tool-Calling Agents on Consumer Hardware with LFM2-24B-A2B" on 02.24.2026. These are first-party creator milestones est The same index also lists later LFM2.5 family variants (350M, 1.2B, 2.6B, VL models, etc.), situating LFM2-24B-A2B within Liquid AI's broader on-device LFM roadmap and showing it preceded the LFM2.5 generation. This positions LFM2-24B-A2B as the largest "scaling-up" member of the LFM2 family prior to the LFM2.5 refresh

The Opper AI release tracker lists LFM2 24B A2B as one of ten dated Liquid AI model releases, confirming it shipped on February 25, 2026 as the largest entry in the LFM2 family at that point in the timeline. The page places the model within Liquid AI's broader release cadence — following LFM2 1.2B (July 10, 2025), LFM2 By positioning LFM2 24B A2B inside a sortable, newest-first Liquid AI release timeline, the tracker provides a timestamped reference point that establishes when this 24B-parameter MoE model entered the Liquid AI lineup. This timeline framing is useful for distinguishing LFM2 24B A2B from its predecessor models in the L