Spark-X2.6-4B: On-Device Specs (Not Meta Muse)

Spark-X2.6-4B is an ungated ~4B text model from XHToken (citation on the card: SparkLLM Team). It is not Muse Spark Open — Meta’s gated 816B/42B multimodal MoE. Official hosts we opened on 2026-10-26: the card XHToken/Spark-X2.6-4B (created 2026-10-19T03:19:51Z; last modified 2026-10-21T08:51:53Z; 514 likes / 3,488 downloads), the Base sibling XHToken/Spark-X2.6-4B-Base (29 likes / 348 downloads), the Apache 2.0 LICENSE, XHToken/Spark-X2.6 (845 stars / 117 forks; default branch X2.6), and the Ollama library tag SparkLLM/Spark-X2.6-4B.

This page hangs off ChatGPT alternatives, closed LLM APIs, Claude Code alternatives, and Cursor alternatives. That mapping is editorial. We have no sourced quality comparison against ChatGPT, Claude Code, or Cursor. Same-site on-device peer: Spark-X2.6-4B vs LFM2.5-24B-A2B.

The card title is “Advancing the Frontier of On-Device Agents for Real-World Tasks.” Hugging Face pipeline_tag is text-generation. Languages on the card: en, zh. The card says Spark-X2.6-4B keeps the Spark-X2.5-4B architecture and gains capability through post-training (coding, agent, long-context, multilingual, reasoning). Those adjectives stay vendor.

What we verified (2026-10-26)

Field	Value	Source
Host	huggingface.co/XHToken/Spark-X2.6-4B (not gated)	HF API
Created / last modified	2026-10-19T03:19:51Z / 2026-10-21T08:51:53Z	same
Likes / downloads	514 / 3,488	same
License tag + file	apache-2.0 / Apache License 2.0	HF cardData + LICENSE
Pipeline tag	text-generation	HF API
Architecture	spark2_5 / Spark2_5ForCausalLM	config.json
Params (card)	4B	card
safetensors.parameters	4,112,079,360	HF API
usedStorage (BF16)	12,686,523,518 bytes (~11.8 GiB)	same
Layers / hidden / heads	36 / 2560 / 16 Q, 4 KV	config.json
Vocab	131,072	config.json
Documented context	1,048,576 tokens	max_position_embeddings + SGLang --context-length
Thinking levels (card)	none, low, high	card
Base sibling	Spark-X2.6-4B-Base · 29 likes · apache-2.0	HF API
GitHub	XHToken/Spark-X2.6 · 845 stars / 117 forks · Apache-2.0	GitHub API

How to run (from the card, not tested)

Copied names only. forkalt has not loaded the checkpoint.

Transformers (card: pip install transformers==4.57.1; trust_remote_code=True; XHToken/Spark-X2.6-4B). Chat template flag: enable_thinking=True.

Ollama (card: v0.34.1 or later):

ollama run SparkLLM/Spark-X2.6-4B

llama.cpp (card: native spark2_5 needs b10828 or later) on a GGUF path. LM Studio 0.4.0+ is named (search “Spark-X2.6”). Spark-MLX-LLM runs the Hugging Face checkpoint on Apple silicon / Linux CPU / CUDA without a GGUF conversion (XHToken/Spark-MLX-LLM).

SGLang and vLLM recipes on the card use the shipped chat_template.jinja, --tool-call-parser spark25, and --reasoning-parser qwen3. The SGLang example sets --context-length 1048576 and says to lower it when memory is tight. Fine-tune path the card names: Llama-Factory.

License

LICENSE is Apache License 2.0. Card tag apache-2.0. Same tag on the Base sibling. GitHub license API also returns Apache-2.0. Not legal advice.

What we did not invent

• Card lines about “state-of-the-art among similarly sized open-source models” and matching ~9B models stay vendor.
• Official SVG figures (images/benchmark.svg, images/all_benchmark.svg, images/inference.svg) are self-reported. We do not copy tok/s or bench scores from those plots into a ranking.
• Peak VRAM and laptop-class tokens/s are not independently measured.
• We do not rank Spark-X2.6-4B against ChatGPT, LFM2.5-24B-A2B, Flint, or Muse Spark Open.

Related on-device / compact cards: LFM2.5-24B-A2B — Spark-X2.6-4B vs LFM2.5-24B-A2B — Flint (39.8B/2.5B coding MoE), Raon-1-mini (28.8B/4.3B preview), MiniCPM-V 4.7 (on-device VLM). Serving: vLLM. Agents: Cline, OpenCursor. Catalog home: forkalt. Editorial: About. Pixel editors are a different search: PhotoCraft vs GIMP.