Spark-X2.5: Two Small Open Models With a One Million Token Window

Two sizes, one architecture, a very long context

The Spark-X2.5 series is introduced as two compact, general-purpose language models, Spark-X2.5-4B and Spark-X2.5-1.7B, positioned as making capable AI more practical and accessible rather than maximal. The stated scope of tasks is broad: conversation, writing, translation, reasoning, coding, tool use and agentic workflows. The headline architectural claim is a native context window of up to 1M tokens together with support for more than 200 languages. The way that context length is reached is the interesting engineering detail. The models use a hybrid attention architecture combining one full-attention layer with three sliding-window attention layers, and the stated effect is that it substantially reduces the computational overhead typically associated with long-context models while natively supporting the long window. Full attention on a subset of layers keeps global context available, while the sliding-window layers do the bulk of the work locally.

The repository is a model card, not a codebase

It is worth being clear about what this repository actually is. The top level contains a `.gitignore`, a code of conduct, the licence, the readme, a security policy and an `assets/` directory, and nothing else. There is no source tree, no training code, no evaluation harness and no serving implementation. The models themselves are distributed as downloadable weights through hosting platforms rather than built from this repository. The repository publishes no GitHub releases, and the last push was on 2026-10-19, so the readme itself is actively maintained while the weights live elsewhere. The project homepage is at xhtoken.ai, and the topic tags point at the ecosystem the models are meant to run in, including llama.cpp, Ollama, vLLM, SGLang, MLX and MLX-LM. For an evaluation, that means the question is what the weights do and which runtimes accept them, not what the code does.

Where to get the weights, by region and by hardware

The download table is organised by platform, and the recommendation depends on where you are and what you run. The official Hugging Face collection is the primary source. ModelScope and AtomGit are both recommended for users in China, Modelers is recommended for users with Ascend chips, and SCNet is recommended for users with Hygon chips. Ollama is listed as the route to download and run the model locally. A note underneath advises that if Hugging Face is slow or unavailable in your region, you should try ModelScope, Modelers, SCNet or AtomGit instead. That is a practical constraint worth taking seriously, because a model distribution that is unreachable from your network is not a model you can deploy, and the choice of mirror is dictated by geography rather than by preference.

Quantised builds released alongside the full weights

On 2026-09-04 the project added FP8 and INT8 quantized versions of both Spark-X2.5-4B and Spark-X2.5-1.7B. That timing matters: the quantized variants appeared two days before the global launch across Hugging Face, ModelScope, Ollama, Modelers and SCNet on 2026-09-01, and one day after deployment support for vLLM and SGLang on Ascend NPU. The sequence suggests the intended path is local or on-device inference rather than a large server deployment, since a 1.7B and a 4B model at 8-bit precision are the sizes that fit on a workstation, let alone a laptop. The description of the series as pushing the limits of agentic capabilities in on-device models points the same way, and the fine-tuning frameworks named as supported, LLaMA-Factory among them, are the usual route for adapting a base checkpoint to a specific task.

Framework support arrived in a scattered order

The release log is a dated list of integrations, and the dates matter for anyone planning an upgrade path. On 2026-09-06 native model architecture support for Spark-X2.5, under the name Spark2_5ForCausalLM, landed in llama.cpp at release b10829. LM Studio support followed on 2026-09-08, and Ollama support on 2026-09-15 for the Spark-X2.5 model architecture. On 2026-09-11 the model got native support in the PocketPal AI app on the App Store and Google Play, and on 2026-09-03 deployment support for vLLM and SGLang on Ascend NPU was added. The model series itself launched globally on 2026-09-01, and AtomGit was added as a distribution channel on 2026-09-02. The practical consequence is that your runtime of choice was supported at a specific moment, and pinning to a release known to include the architecture is the safer choice than tracking the newest build.

Agent harnesses, accelerators and the training story

Two technical highlights are worth separating. The first is integration: the models are described as deeply integrated with popular agent harnesses, including Codex, Claude Code, OpenClaw and Hermes, with state-of-the-art results claimed among models of comparable size across coding, agentic workflows, reasoning and instruction-following. The second is hardware breadth, with NVIDIA, Huawei, Hygon and HOUMO.AI named as supported platforms, and claimed advantages in TTFT, TOPT and overall inference efficiency compared with similarly sized models. The training story is that the models were trained on Huawei Ascend clusters, using large-scale reinforcement learning and post-training techniques, with MOPD specifically named as a contributor to reasoning, coding, agentic and instruction-following capability. The efficiency claims are the vendor's own, and no independent measurement appears in the documentation.

Licence, community channels and the gap in the documentation

The licence is Apache-2.0, which is permissive enough for commercial deployment and modification, and the security policy file sits alongside it, so vulnerability reporting has a defined route. The community channels are wide for a model release: Slack, Discord, a YouTube channel, a dev.to presence, Bluesky, X, Zhihu and a WeChat group, with questions invited through GitHub issues. The gap is in what is written down. The benchmark table in the readme compares the two Spark sizes against other models including a Qwen3.5-9B column, but the surrounding text explains the table in general terms rather than walking through individual results, and no evaluation methodology, prompt set or harness configuration is described. For a model whose selling point is agentic quality on a small footprint, that is the part a careful adopter wants spelled out, and it is not written down here.

Editorial conclusion

Spark-X2.5 fits a team that wants a general-purpose model small enough to run on a workstation or laptop, in an agent harness such as Codex or Claude Code, with a context window long enough to hold a large codebase or long transcript. It is the wrong choice if you need a chat model with an existing quantisation ecosystem and no new architecture support work, and the wrong choice if the documentation gaps matter to you, since this repository holds a model card rather than code. Before adopting it, check the framework build you plan to use, because native architecture support arrived in llama.cpp, Ollama, LM Studio, vLLM and SGLang at different dates.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. Project website
4. README
5. XHToken/Spark-X2.5 on GitHub