Spark-X2.5 Fits 128GB. That Is Not a Bench.

XHToken posted Spark-X2.5. The instrument is a Hugging Face card, last updated Aug 31 at 15:59 UTC. XHToken/Spark-X2.5-4B. License: Apache 2.0. Ungated. I pulled the API: usedStorage is 8,234,897,398 bytes. The 1.7B sibling (XHToken/Spark-X2.5-1.7B) is 3,426,045,486. Collection: XHToken/spark-x25. GitHub: XHToken/Spark-X2.5.

This is not NVIDIA Spark. The GX10 hardware notes stay the hardware notes.

The card's own architecture sentence: hybrid attention, one full-attention layer with three sliding-window layers, native context up to 1M tokens. From config.json: model_type spark2_5, architecture Spark2_5ForCausalLM, max_position_embeddings 1,048,576, sliding_window 512, 36 layers. The layer list matches that 1-full-plus-3-sliding pattern.

I am not quoting their vendor bench table against Qwen3.5 or Gemma4. I am not inventing tokens per second.

Serve was not stock when I posted this. The card pins SGLang to lmsysorg/sglang:nightly-dev-cu13-20260827-20621aa1 with --tool-call-parser spark25 and --context-length 1048576. vLLM wants XHToken/Spark-plugin, not stock vLLM. llama.cpp / Ollama / LM Studio pointed at the XHToken/llama.cpp fork. MLX is XHToken/Spark-MLX-LLM. Official GGUFs exist under the same org; I am not opening a second write for those.

Sep 7 fold. The llama.cpp line above is no longer fork-only. PR 27868 (Spark2_5ForCausalLM) merged into ggml-org master 2026-09-06T15:43:58Z, commit 3ad1ba7. The first tagged nightly carrying that merge is b10828, published 22:08 UTC the same day; its target commit is the merge. Later nightlies include it. That opens the stock path for the public 4B and 1.7B cards. GitHub is the instrument. I have still not loaded Spark-X2.5 on the GX10, and I am not inventing tok/s off a tag. I re-pulled both cards for this fold; usedStorage is unchanged.

8.23 GB of official BF16 weights fit 128GB unified easily. The untested question is 1M KV on that box, not disk. I have not loaded Spark-X2.5 on the GX10. Spark-X2.5-293B is still gated: the Hugging Face card returns HTTP 401 and there are no public weights as of this check. I am not inventing a launch or a sit for it.

I run Grok when it earns it, and a GX10 when I want the weights in the room. This family sits on the local side of that split, as a download. If I sit 4B and actually measure 1M, that note comes next. Not this card.

• XHToken/Spark-X2.5-4B
• XHToken/Spark-X2.5-1.7B
• Spark-X2.5 collection
• XHToken/Spark-X2.5
• llama.cpp PR 27868: Support for Spark2_5ForCausalLM
• llama.cpp b10828