Machine-generated, human-curated. Surfaced by a sourcing pipeline (HuggingFace trending + newest, drilled down to the authoritative model cards) and written by an AI assistant under human curation. Adoption counts come from the HuggingFace API; benchmark figures are the labs’ own unless the post says measured.
A tiny adorable cartoon robot with one spark-shaped antenna flexing tiny arms while lifting a barbell labeled 9B; beside it a taller robot labeled 12B looks shocked as its own weights tumble to the ground.
Small weights, big lift — if the table survives contact with strangers.

Spark-X2.5-4B launched Sep 1 with the pitch every small model wants: a chart where a 4B beats Qwen3.5-9B and Gemma 4 12B on agentic work — τ³-bench 30.4, SWE-Bench Pro 44.4, AIME 2026 90.7. That chart is XHToken’s own, and the model is four days old, so nobody independent has touched it. What I can verify from outside: the official GGUF sits at 10.9k downloads against the base repo’s 3,524, eleven community quant builds exist, MLX builds for Apple Silicon landed within days, official FP8 and INT8 versions followed three days later, and Codex, Claude Code, OpenClaw and Hermes already ship native integrations — including a spark25 tool-call parser in SGLang and vLLM. Stars are applause; loaders are the vote. The vote says yes.

The object itself: XHToken’s SparkLLM team, out of the Huawei Ascend ecosystem, Apache-2.0, roughly 20T tokens of training with RL post-training, 200+ languages claimed, pitched at on-device agentic work — conversation, coding, tool use, full workflows. There’s a 1.7B sibling for anyone who wants less of all of that. It runs on every stack you already have: vLLM, SGLang, llama.cpp, MLX, Ollama, LM Studio. A Q4-class GGUF lands around 2.5–3.5 GB (estimate), so an 8–12 GB Mac or any recent consumer GPU holds it without a conversation.

The genuinely clever part isn’t the size, it’s how it carries a million tokens without eating your RAM. Standard attention scores every token against every token, and for long contexts it’s the cache, not the weights, that bankrupts you — a 1M-token KV cache can ask for tens of GB before an answer arrives. Spark-X2.5 goes hybrid instead: one full-attention layer per three sliding-window layers per block. The windowed layers cap their cache at the window size no matter how long the context runs; the periodic full layer keeps global reasoning. That’s how a 2.5 GB 4-bit model claims a context length that would drown a full-attention 4B outright. The honest caveat, printed in the idea itself: a bounded window can miss a distant token — an early instruction, a detail buried near token 900,000 — and long-context retrieval quality is notoriously hard to verify. The 1M number is the vendor’s; run your own recall test before trusting it with a big trace.

So here’s where I land: the benchmarks are a vendor table and GitHub sits at 151 stars, which is not a stampede — 488 likes in four days is solid interest, not conversion. What makes this worth an evening is the boring trio: same-or-smaller footprint than the Qwen3.5-4B you run today, native 1M context, and zero adapter work in the harnesses you already drive, with Apache-2.0 on top so nobody can rug you on licensing later. Swapping in and comparing costs you one download. The 4B-beats-9B line is a question I’ll let other people answer first; the size-and-context win is available today. The fine-tune bench around it is empty — every derivative sits at zero downloads, which just means brand-new, but nothing derivative is interesting yet. Either way, this is the shape the field is converging on: the question stopped being “which agent is smarter” and became “can you change the thing you’re driving.”

Download, like and build counts verified via the HuggingFace API. The benchmark table, the 200+ language claim and the 1M-context numbers are XHToken’s own; size figures marked estimate are arithmetic on published formats, not measurements.