LFM3 vs LFM2.5-24B-A2B

Official Liquid AI LFM3-1.7B edge-inference figure from the Hugging Face card — LiquidAI/LFM3-1.7B card assets/lfm3-edge-inference.webp. Source · License: Liquid AI product image.

Official Liquid AI LFM2.5-24B-A2B blog Open Graph image — Liquid AI official blog OG. Source · License: Liquid AI product image.

Verdict. Pick LFM3 if you want Liquid AI’s October 2026 ~2B-class family (text 1.73B / 256K, plus VL, Audio, Omni, and d1 cards) on the new tri-hybrid backbone. Pick LFM2.5-24B-A2B if you want the existing 24.0B / 2.45B text MoE already documented here. Both use LFM Open License v1.0 (USD 10 million commercial-use threshold). They are different generations. Vendor benches stay vendor. We have no sourced bake-off.

Compare LFM3 and LFM2.5-24B-A2B.

Dimension	LFM3	LFM2.5-24B-A2B	Source
Generation	LFM3 family (blog 27 Oct 2026)	LFM2.5 text MoE (card 12 Oct 2026)	Official blog / HF createdAt
License	LFM Open License v1.0 (USD 10M commercial-use threshold)	LFM Open License v1.0 (USD 10M commercial-use threshold)	Hugging Face LICENSE files (2026-10-28 / 2026-10-14)
Gating	No (all family cards we opened)	No	Hugging Face API
Developer	Liquid AI	Liquid AI	Official cards
Parameters (lead card)	1.73B text (safetensors 1,728,801,536)	24.0B total / 2.45B active	Official READMEs / HF API
Architecture	Tri-hybrid 24 layers (13 conv + 5 SWA + 6 global)	LFM2.5 MoE (40 layers; 64 experts, top-4)	Official READMEs / blog
Modalities	Family covers text, image, video, audio, speech, decision	text-generation only	Official cards
Context length (lead)	256K (LFM3-1.7B)	131,072	Official READMEs
Hugging Face likes (lead)	160 (LFM3-1.7B, 2026-10-28)	170 (2026-10-14)	Hugging Face API
Quality vs ChatGPT / Claude	Not independently tested	Not independently tested	n/a

When to choose which

~2B on-device family with VL / audio / omni siblings: LFM3 — Family cards we opened. Not a quality rank.
24B/2.45B text MoE already documented on this site: LFM2.5-24B-A2B — Identification of size class, not a bake-off.
256K context on the text card: LFM3 — LFM3-1.7B README is 256K. LFM2.5 README is 131,072. Untested here.
Same LFM Open License v1.0 grant: Tie — Both LICENSE files are v1.0 with the USD 10M threshold. Not legal advice.

Read the entity pages: LFM3 and LFM2.5-24B-A2B. Both hang off closed LLM APIs. LFM3 also hangs off ChatGPT alternatives. LFM3 ≠ LFM2.5.

Comparison graphic: LFM3 ~2B multimodal family versus LFM2.5-24B-A2B 24B/2.45B text MoE under the same LFM Open License v1.0 — forkalt comparison graphic (2026-10-28). Not a ranking.

Why this comparison exists

Searchers who already have the LFM2.5-24B-A2B page will hit LFM3 as a same-brand name. They are not the same checkpoint. LFM3 is the 27 October 2026 ~2B family (text / VL / audio / omni / d1). LFM2.5-24B-A2B is the 24B/2.45B text MoE (card created 12 October 2026). This is identification, not a bake-off.

Sourced snapshot (2026-10-28 / 2026-10-14)

Field	LFM3 (lead: LFM3-1.7B)	LFM2.5-24B-A2B
Host	huggingface.co/LiquidAI/LFM3-1.7B (not gated)	huggingface.co/LiquidAI/LFM2.5-24B-A2B (not gated)
HF likes / downloads	160 / 6,311 (2026-10-28)	170 / 2,565 (2026-10-14)
License	LFM Open License v1.0	LFM Open License v1.0
Size	1.73B text; family ~2B	24.0B / 2.45B
Job	Family: text + VL + audio + omni + decision	text-generation
Context	256K (text card)	131,072
Official local path	GGUF / MLX / ONNX / FP8 named	GGUF / MLX / ONNX / llama.cpp
Quality vs ChatGPT	Not independently tested	Not independently tested

License

Both LICENSE files are LFM Open License v1.0. Commercial Use is conditioned on You/Legal Entity not exceeding USD 10,000,000 annual revenue. Above that threshold, Commercial Use is not licensed under that agreement. Qualified non-profits for non-commercial/research use are exempt from the threshold. Not legal advice.

How to choose

Need the October 2026 ~2B family (VL/audio/omni/d1): LFM3. Dedicated text card: LFM3-1.7B. Dedicated omni card: LFM3-Omni-2B. Sibling compare: LFM3-1.7B vs LFM3-Omni-2B.
Need Liquid’s 24B/2.45B text MoE: LFM2.5-24B-A2B.
Need Apache-2.0 ~4B dense with an Ollama one-liner: Spark-X2.6-4B — Spark vs LFM2.5.
Need Apache-2.0 on-device VLM: MiniCPM-V 4.7.
Engine page: vLLM.

See also AI Chat & Agents and Models. Catalog: home. Editorial: About.

FAQ

Which is better at on-device agents?

We do not rank them. LFM3-1.7B’s own README says it is not recommended for agentic coding or complex multi-step reasoning because of its small size. LFM2.5-24B-A2B’s blog tables stay vendor. See the entity pages.

Are these the same weights?

No. LFM3 is the 27 October 2026 family. LFM2.5-24B-A2B is the earlier 24B/2.45B MoE (card created 12 October 2026). Same licensor and license family; different generation.

Why this pairing?

LFM2.5-24B-A2B is the Liquid page already on this site. The task is to keep LFM3 distinct from that checkpoint. Spark-X2.6-4B remains a separate Apache-2.0 ~4B compare against LFM2.5.

Is LFM3 d1 the same as LiquidAI/d1-3B?

No. LFM3 ships d1-2B and d1-Omni-2B under LFM Open License v1.0. LiquidAI/d1-3B (LFM2.5-VL-3B fine-tune) stays on the classification watch list with no dedicated page.

Sources & verification

Last verified 2026-10-28 · Last updated 2026-10-28· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.