Spark-X2.6-4B vs LFM2.5-24B-A2B

Official Spark-X2.6-4B edge-inference figure from the XHToken Hugging Face card — XHToken/Spark-X2.6-4B card images/inference.svg. Source · License: Apache-2.0.

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

Verdict. Pick Spark-X2.6-4B if you want XHToken’s ungated Apache-2.0 ~4B dense on-device agent with a LICENSE file, 1M context in config.json, and first-party Ollama / llama.cpp / Transformers names. Pick LFM2.5-24B-A2B if you want Liquid AI’s 24.0B/2.45B on-device MoE with GGUF/MLX/ONNX siblings and can accept LFM Open License v1.0 (USD 10 million commercial-use threshold). Vendor benches stay vendor. We have no sourced bake-off. Qwen3.5-4B is not a page on this site.

Compare Spark-X2.6-4B and LFM2.5-24B-A2B.

Dimension	Spark-X2.6-4B	LFM2.5-24B-A2B	Source
License	Apache-2.0 (LICENSE file + card tag)	LFM Open License v1.0 (USD 10M commercial-use threshold)	Hugging Face cards / LICENSE files (2026-10-26 / 2026-10-14)
Gating	No	No	Hugging Face API
Developer	XHToken (SparkLLM Team)	Liquid AI	Official cards
Parameters	4B on the card; HF safetensors 4,112,079,360	24.0B total / 2.45B active	Official cards / HF API
Architecture	Dense spark2_5 / Spark2_5ForCausalLM (36 layers)	LFM2.5 MoE (40 layers; 64 experts, top-4)	Official configs / README
Modalities	text-generation	text-generation	Hugging Face API
Context length	1,048,576 (config.json + SGLang recipe)	131,072	Official cards / configs
Languages (card)	en, zh	EN, AR, ZH, FR, DE, IT, JA, KO, PT, ES, VI, TH, ID, HI, RU, PL	Official cards
Serving named	Transformers 4.57.1, SGLang, vLLM, Spark-MLX-LLM, llama.cpp b10828+, Ollama, LM Studio	llama.cpp, MLX, vLLM, SGLang, ONNX (blog / README)	Official cards
Hugging Face likes	514 (2026-10-26)	170 (2026-10-14)	Hugging Face API
Quality vs ChatGPT / Claude	Not independently tested	Not independently tested	n/a

When to choose which

Apache-2.0 with a LICENSE blob and no revenue threshold: Spark-X2.6-4B — Identification of the grant, not legal advice.
Smaller dense ~4B checkpoint: Spark-X2.6-4B — Card size 4B vs Liquid 24.0B/2.45B. Not a quality rank.
Published 2.45B-active MoE with GGUF/MLX/ONNX siblings on day one: LFM2.5-24B-A2B — Liquid README lists those siblings. Spark names GGUF via llama.cpp / community paths.
1M-token config figure: Spark-X2.6-4B — Spark config.json is 1,048,576. LFM README is 131,072. Untested here.

Read the entity pages: Spark-X2.6-4B and LFM2.5-24B-A2B. Spark hangs off ChatGPT alternatives, closed LLM APIs, Claude Code alternatives, and Cursor alternatives. LFM hangs off the closed-API hub. Spark is not Muse Spark Open.

Comparison graphic: Apache-2.0 Spark-X2.6-4B ~4B dense versus LFM2.5-24B-A2B 24B/2.45B MoE under LFM Open License v1.0 — forkalt comparison graphic (2026-10-26). Not a ranking.

Why this comparison exists

Searchers who want a small on-device open agent now have XHToken’s ~4B dense Spark-X2.6-4B (card created 2026-10-19) next to Liquid AI’s 24B/2.45B LFM2.5-24B-A2B (card created 2026-10-12), which is the on-device-agent page already on this site. Spark’s own card names Qwen3.5-4B and Gemma-E4B as size-class peers; those are not pages here. This is identification, not a bake-off.

Sourced snapshot (2026-10-26 / 2026-10-14)

Field	Spark-X2.6-4B	LFM2.5-24B-A2B
Host	huggingface.co/XHToken/Spark-X2.6-4B (not gated)	huggingface.co/LiquidAI/LFM2.5-24B-A2B (not gated)
HF likes / downloads	514 / 3,488 (2026-10-26)	170 / 2,565 (2026-10-14)
License	Apache-2.0 (LICENSE file)	LFM Open License v1.0
Size	4B (card); ~4.11B safetensors	24.0B / 2.45B
Job	text-generation; on-device agent/coding	text-generation; on-device agent
Context	1,048,576	131,072
Official local path	Ollama / llama.cpp / Transformers	GGUF / MLX / ONNX / llama.cpp
Quality vs ChatGPT	Not independently tested	Not independently tested

License

Spark’s LICENSE file is Apache License 2.0. Liquid’s LICENSE is 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 Apache-2.0 ~4B dense with an Ollama one-liner: Spark-X2.6-4B.
Need Liquid’s 24B/2.45B MoE and can accept LFM 1.0: LFM2.5-24B-A2B.
Need Apache-2.0 39.8B/2.5B local coding MoE: Flint — Flint vs Ornith-1.5-122B.
Need a 28.8B/4.3B compact preview MoE: Raon-1-mini.
Need Meta’s gated 816B/42B multimodal MoE: Muse Spark Open.
Engine page: vLLM.

See also AI Chat & Agents and Dev & Coding. Catalog: home. Editorial: About. Pixel editors are a different search: PhotoCraft vs GIMP.

FAQ

Which is better at on-device agents?

We do not rank them. Both cards use “on-device agent” language. Spark’s inference and benchmark SVGs and Liquid’s LiveClaw / AutomationBench table stay vendor. See the entity pages.

Which is easier to run with Ollama?

Spark’s card pastes ollama run SparkLLM/Spark-X2.6-4B (v0.34.1+). Liquid’s README lists GGUF for llama.cpp and does not paste that Ollama one-liner on the check we used. Untested here.

Why this pairing?

There is no Qwen3.5-4B or Gemma-E4B page on this site. LFM2.5-24B-A2B is the existing on-device-agent card. Flint is a larger 39.8B/2.5B coding MoE; Raon-1-mini is a 28.8B/4.3B preview. Those stay related links.

Is Spark-X2.6-4B Meta Muse Spark?

No. Muse Spark Open is Meta’s gated 816B/42B multimodal MoE. See Muse Spark Open.

Sources & verification

Last verified 2026-10-26 · Last updated 2026-10-26· Edited by forkalt editors · Editorial policy

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