Spark-X2.5

XHToken released the Spark-X2.5 open model series, including Spark-X2.5-4B and Spark-X2.5-1.7B, two compact general-purpose language models with native context windows of up to 1M tokens and support for more than 200 languages. The models use a hybrid attention architecture combining one full-attention layer with three sliding-window attention layers, were trained on Huawei Ascend clusters, and are compatible with vLLM, SGLang, llama.cpp, MLX, Ollama, LM Studio and LLaMA-Factory across NVIDIA, Huawei, Hygon and HOUMO.AI hardware. The global launch on Hugging Face, ModelScope, Ollama, Modelers and SCNet was followed by FP8 and INT8 quantized versions on 2026/09/04 and native llama.cpp support (Spark2_5ForCausalLM) on 2026/09/06.

Welcome to the GitHub repository of Spark-X2.5 open model series. You can find official information about Spark-X2.5, and post your questions here Issues https://github.com/XHToken/Spark-X2.5/issues . Today, we are introducing Spark-X2.5-4B and Spark-X2.5-1.7B, two compact, general-purpose language models designed to make capable AI more practical, efficient, and accessible. The models deliver strong performance across a broad range of everyday tasks—including conversation, writing, translation, reasoning, coding, tool use, and agentic workflows—achieving leading results among open-source models of comparable size. Spark-X2.5 combines an efficiency-oriented architecture with native context windows of up to 1M tokens, and support for more than 200 languages. Technical Highlights : - Efficient Architecture and Native 1M-token Context : The models use a hybrid attention architecture that combines one full-attention layer with three sliding-window attention layers. This design substantially reduces the computational overhead typically associated with long-context models while natively supporting a context window of up to 1M tokens. - Strong Coding and Agent Capabilities : The models are deeply integrated with popular agent harnesses, including Codex, Claude Code, OpenClaw, and Hermes. They deliver state-of-the-art performance among models of comparable size across everyday coding, agentic workflows, reasoning, and instruction-following tasks. - Broad Hardware and Software Compatibility : The models support a wide range of hardware platforms, including NVIDIA, Huawei, Hygon, HOUMO.AI, etc. It is compatible with leading inference frameworks such as vLLM, SGLang, llama.cpp, MLX, and can be deployed quickly through platforms including Ollama and LM Studio. The models can also be customized using popular fine-tuning frameworks such as LLaMA-Factory. Across multiple hardware platforms, they deliver superior TTFT, TOPT, and overall inference efficiency compared with similarly sized models. - Advanced Training Algorithms : The models were trained on Huawei Ascend clusters. Large-scale reinforcement learning and post-training techniques such as MOPD significantly enhance its reasoning, coding, agentic, and instruction-following capabilities. - 2026/09/15 🤝 Added ollama https://github.com/ollama/ollama/releases/tag/v0.34.1 support for Spark-X2.5 model architecture. - 2026/09/11 🚀 Added native support for Spark-X2.5 model in PocketPal AI app, now available on App Store https://apps.apple.com/us/app/pocketpal-ai/id6502579498 and Google Play https://play.google.com/store/apps/details?id=com.pocketpalai . - 2026/09/08 🤝 Added LM Studio https://lmstudio.ai/download support for Spark-X2.5 model architecture. - 2026/09/06 🤝 Added native model architecture support for Spark‑X2.5 Spark2 5ForCausalLM in llama.cpp https://github.com/ggml-org/llama.cpp/releases/tag/b10829 . - 2026/09/04 🚀 Released FP8 and INT8 quantized versions of Spark-X2.5-4B and Spark-X2.5-1.7B. - 2026/09/03 🤝 Added deployment support for vLLM and SGLang on Ascend NPU. - 2026/09/02 🚀 Added AtomGit https://ai.atomgit.com/collections/2095030878254981121 as a new distribution channel. - 2026/09/01 🚀 Global launch of the Spark-X2.5 model series on Hugging Face https://huggingface.co/collections/XHToken/spark-x25 , ModelScope https://www.modelscope.cn/collections/XHToken/Spark-X25 , Ollama https://ollama.com/SparkLLM , Modelers https://modelers.cn/user/XHToken , and SCNet https://www.scnet.cn/ui/aihub/models/XHToken/Spark-X2.5-4B . The Spark-X2.5-4B and Spark-X2.5-1.7B are available on the following platforms. Choose the most suitable download source for your region and environment: | Platform | Download | Description | |---|---|---| | Hugging Face | Spark-X2.5 in Huggingface https://huggingface.co/collections/XHToken/spark-x25 | Official Hugging Face model collection for Spark-X2.5 | | ModelScope | Spark-X2.5 in ModelScope https://www.modelscope.cn/collections/XHToken/Spark-X25 | Recommended download source for users in China | | Modelers | Spark-X2.5 in Modelers https://modelers.cn/user/XHToken | Recommended download source for users with Ascend chips | | Ollama | Spark-X2.5 in Ollama https://ollama.com/SparkLLM | Download and run the model locally with Ollama | | SCNet | Spark-X2.5 in SCNet https://www.scnet.cn/ui/aihub/models/XHToken/Spark-X2.5-4B | Recommended download source for users with Hygon chips | | AtomGit | Spark-X2.5 in AtomGit https://ai.atomgit.com/collections/2095030878254981121 | Recommended download source for users in China | If Hugging Face is slow or unavailable in your region, try ModelScope, Modelers, SCNet or AtomGit instead. | Benchmark | Spark‑X2.5‑4B | Spark‑X2.5‑1.7B | Qwen3.5‑9B | Qwen3.5‑4B | Qwen3.5‑2B | Gemma4‑12B | Gemma4‑E4B | Gemma4‑E2B | |---|---|---|---|---|---|---|---|---| | Agent | | | | | | | | | | BFCL‑V4 | 65.1 | 46.9 | 66.1 | 50.3 | 43.6 | 37.4 | 36.9 | 30.2 | | τ²‑bench | 75.1 | 65.3 | 79.1 | 79.9 | 48.8 | 69.0 | 42.2 | 24.5 | | τ³‑bench | 30.4 | 20.1 | 9.3 | 6.7 | 4.1 | 13.3 | 10.1 | 8.8 | | MCP‑Atlas | 54.6 | 23.4 | 47.4 | 40.8 | 14.8 | 30.5 | 15.0 | 12.6 | | MCP‑Mark | 14.2 | 2.3 | 13.4 | 12.5 | – | – | – | – | | Workspace Bench | 31.2 | 18.9 | 25.5 | 21.3 | 7.7 | – | – | – | | VitaBench2.0 | 25.2 | 8.3 | 15.6 | 18.2 | 5.2 | 12.4 | 4.8 | 4.4 | | BrowseComp | 40.9 | 29.7 | 8.3 | 14.3 | 3.1 | 10.0 | 8.3 | 3.7 | | Code | | | | | | | | | | SWE‑Bench Pro | 44.4 | 10.4 | 33.8 | 29.4 | 1.9 | 21.9 | 4.0 | – | | SWE‑Bench Verified | 41.6 | 28.3 | 53.1 | 38.8 | 6.8 | 44.2 | 14.0 | – | | SWE‑Bench Multilingual | 53.3 | 23.3 | 43.3 | 27.7 | 5.0 | 32.5 | – | – | | SciCode | 34.7 | 18.2 | 32.7 | 24.0 | 6.0 | 39.8 | 27.5 | 20.5 | | Math | | | | | | | | | | Gaokao 2026 | 133.4 | 114.8 | 135.5 | 130.3 | 94.0 | 130.6 | 102.4 | 81.8 | | AIME 2026 | 90.7 | 69.4 | 88.2 | 83.0 | 30.8 | 82.1 | 42.5 | 37.5 | | HMMT Feb 2026 | 81.2 | 48.4 | 70.8 | 69.7 | 21.5 | 65.6 | 34.2 | 20.5 | | IMO‑AnswerBench | 74.2 | 45.4 | 69.8 | 68.5 | – | 57.2 | 26.9 | 22.6 | | General & Knowledge | | | | | | | | | | IFEval | 93.0 | 89.5 | 91.5 | 89.8 | 78.6 | 94.8 | 45.3 | 34.8 | | IFBench | 75.0 | 66.3 | 64.5 | 59.2 | 41.3 | 73.5 | 44.0 | 22.7 | | AA‑LCR | 56.3 | 24.3 | 63.0 | 57.0 | 25.6 | 55.3 | 34.7 | 18.3 | | HLE | 12.3 | 6.3 | 14.3 | 8.6 | 2.1 | 13.1 | 3.9 | 2.5 | | GPQA | 67.4 | 43.8 | 77.2 | 67.2 | 44.6 | 72.8 | 54.5 | 43.8 | - denotes reported results from publicly‑released model cards / papers and - denotes scores not yet available. - All evaluations are conducted in thinking mode. The recommended sampling parameters for Spark-X2.5 are temperature=1.0, top p=0.95, and top k=-1. - Gaokao 2026 consists of the five 2026 Chinese GAOKAO examinations National I,National II, Beijing, Shanghai, Tianjin , each graded out of 150 points. The examples below serve a local Spark-X2.5-4B checkpoint. Set MODEL PATH to its absolute path before starting a container: export MODEL PATH=/absolute/path/to/Spark-X2.5-4B Use the pre-built image that tracks the Spark-X2.5 runtime: docker pull lmsysorg/sglang:nightly-dev-cu13-20260827-20621aa1 A3 daily build export SGLANG IMAGE=quay.io/ascend/sglang:main-cann9.0.0-a3 A2 daily build use this instead on A2 hardware export SGLANG IMAGE=quay.io/ascend/sglang:main-cann9.0.0-910b docker pull "$SGLANG IMAGE" The following commands start an OpenAI-compatible API server configured for a maximum context length of 1,048,576 tokens. This setting requires sufficient device memory; reduce --context-length when necessary. docker run --rm -it \ --gpus '"device=0"' \ --ipc=host \ -p 30000:30000 \ -v "$MODEL PATH":/root/Spark-X2.5-4B:ro \ lmsysorg/sglang:nightly-dev-cu13-20260827-20621aa1 \ python -m sglang.launch server \ --model-path /root/Spark-X2.5-4B \ --served-model-name spark2.5 \ --tool-call-parser spark25 \ --reasoning-parser qwen3 \ --tp-size 1 \ --mem-fraction-static 0.8 \ --context-length 1048576 \ --chat-template /root/Spark-X2.5-4B/chat template.jinja \ --host 0.0.0.0 \ --port 30000 docker run -it --rm -e ASCEND USE FIA=1 --network=host --ipc=host --shm-size=16g \ --device=/dev/davinci0 --device=/dev/davinci1 --device=/dev/davinci2 --device=/dev/davinci3 \ --device=/dev/davinci4 --device=/dev/davinci5 --device=/dev/davinci6 --device=/dev/davinci7 \ --device=/dev/davinci8 --device=/dev/davinci9 --device=/dev/davinci10 --device=/dev/davinci11 \ --device=/dev/davinci12 --device=/dev/davinci13 --device=/dev/davinci14 --device=/dev/davinci15 \ --device=/dev/davinci manager \ --device=/dev/devmm svm \ --device=/dev/hisi hdc \ --volume /usr/local/sbin:/usr/local/sbin \ --volume /usr/local/Ascend/driver:/usr/local/Ascend/driver \ --volume /usr/local/Ascend/firmware:/usr/local/Ascend/firmware \ --volume /etc/ascend install.info:/etc/ascend install.info \ --volume /var/queue schedule:/var/queue schedule \ --volume ~/.cache/:/root/.cache/ \ --volume "$MODEL PATH:/root/Spark-X2.5-4B:ro" \ --entrypoint=python \ "$SGLANG IMAGE" \ -m sglang.launch server \ --model-path /root/Spark-X2.5-4B \ --served-model-name spark2.5 \ --tool-call-parser spark25 \ --reasoning-parser qwen3 \ --tp-size 1 \ --mem-fraction-static 0.8 \ --context-length 1048576 \ --chat-template /root/Spark-X2.5-4B/chat template.jinja \ --host 0.0.0.0 \ --port 30000 Thinking is enabled by default by both the chat template and the Qwen3 reasoning parser. To disable thinking for a specific request, set "chat template kwargs": {"enable thinking": false} . curl -s http://localhost:30000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "spark2.5", "messages": { "role": "user", "content": "安徽的省会在哪里？" } , "max tokens": 131072, "temperature": 1, "top k": -1, "top p": 0.95, "repetition penalty": 1, "presence penalty": 0, "frequency penalty": 0 }' vLLM provides an official Docker image for NVIDIA GPU deployment: docker run --rm --gpus all \ --ipc=host \ -p 30000:30000 \ -v "$MODEL PATH:/models/Spark-X2.5-4B:ro" \ vllm/vllm-openai:latest \ --model /models/Spark-X2.5-4B \ --port 30000 \ --trust-remote-code \ --served-model-name spark25 \ --tensor-parallel-size 1 \ --gpu-memory-utilization 0.7 \ --enable-prefix-caching \ --chat-template /models/Spark-X2.5-4B/chat template.jinja For Ascend NPUs, choose an official image for the fastest setup. export IMAGE=quay.io/ascend/vllm-ascend:nightly-main docker pull "$IMAGE" export DEVICE=/dev/davinci0 export MODEL CACHE="${HOME}/.cache" mkdir -p "$MODEL CACHE" docker run --rm \ --name vllm-ascend \ --shm-size=1g \ --device "$DEVICE" \ --device /dev/davinci manager \ --device /dev/devmm svm \ --device /dev/hisi hdc \ -v /usr/local/dcmi:/usr/local/dcmi \ -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \ -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \ -v /etc/ascend install.info:/etc/ascend install.info \ -v "$MODEL CACHE:/root/.cache" \ -p 8000:8000 \ -it "$IMAGE" bash export IMAGE=quay.io/ascend/vllm-ascend:nightly-main-a3 docker pull "$IMAGE" export DEVICE0=/dev/davinci0 export DEVICE1=/dev/davinci1 export MODEL CACHE="${HOME}/.cache" mkdir -p "$MODEL CACHE" docker run --rm \ --name vllm-ascend \ --shm-size=1g \ --device "$DEVICE0" \ --device "$DEVICE1" \ --device /dev/davinci manager \ --device /dev/devmm svm \ --device /dev/hisi hdc \ -v /usr/local/dcmi:/usr/local/dcmi \ -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \ -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \ -v /etc/ascend install.info:/etc/ascend install.info \ -v "$MODEL CACHE:/root/.cache" \ -p 8000:8000 \ -it "$IMAGE" bash export IMAGE=quay.io/ascend/vllm-ascend:nightly-main-a5 docker pull "$IMAGE" export MODEL CACHE="${HOME}/.cache" mkdir -p "$MODEL CACHE" docker run --rm \ --name vllm-ascend \ --net=host \ --shm-size=1g \ --device /dev/davinci0 \ --device /dev/davinci manager \ --device /dev/devmm svm \ --device /dev/hisi hdc \ -v /usr/local/dcmi:/usr/local/dcmi \ -v /usr/local/Ascend/driver/tools/hccn tool:/usr/local/Ascend/driver/tools/hccn tool \ -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \ -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \ -v /etc/ascend install.info:/etc/ascend install.info \ -v "$MODEL CACHE:/root/.cache" \ -it "$IMAGE" bash Install the Spark plugin inside the container: pip install uv uv venv ~/spark2 5 source ~/spark2 5/bin/activate git clone https://github.com/XHToken/Spark-plugin.git cd ./Spark-plugin uv pip install . vllm serve "/models/Spark-X2.5-4B" \ --port "30000" \ --trust-remote-code \ --served-model-name spark25 \ --tensor-parallel-size 1 \ --gpu-memory-utilization 0.7 \ --enable-prefix-caching \ --chat-template /models/Spark-X2.5-4B/chat template.jinja curl -s http://127.0.0.1:30000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "spark25", "messages": {"role": "user", "content": "安徽的省会在哪里？"} , "temperature": 1.0, "top k": -1, "top p": 0.95 }' Spark-MLX-LLM runs the original Spark-X2.5 Hugging Face checkpoints locally. It supports Apple silicon GPU, Linux CPU, and NVIDIA CUDA on Linux. No GGUF conversion is required. git clone https://github.com/XHToken/Spark-MLX-LLM.git cd Spark-MLX-LLM python3 -m venv .venv source .venv/bin/activate Apple silicon python -m pip install -e . Linux CPU python -m pip install -e '. cpu ' Linux with CUDA 12 python -m pip install -e '. cuda12 ' Linux with CUDA 13 python -m pip install -e '. cuda13 ' spark-mlx-generate \ --device gpu \ --dtype bfloat16 \ --model XHToken/Spark-X2.5-1.7B \ --prompt "安徽的省会在哪里？" \ --max-tokens 512 \ --temp 0 1. Download and install Ollama from ollama.com https://ollama.com/download v0.34.1 or later . 2. Run the model: ollama run SparkLLM/Spark-X2.5-1.7B 1. Download and install LM Studio from lmstudio.ai https://lmstudio.ai/download 0.4.0 or later . 2. Search for "Spark-X2.5" in LM Studio and download the model. 3. Load the model and start chatting. List available models lms ls Run chat with Spark-X2.5 lms chat spark-x2.5 Note The latest Unsloth Studio supports native Spark-X2.5 GGUF inference. We recommend using Llama-Factory https://github.com/XHToken/LlamaFactory to fine-tune the model. The Spark-X2.5 model series is licensed under the Apache 2.0 License https://github.com/XHToken/Spark-X2.5/blob/main/LICENSE . If you find our work helpful, feel free to give us a cite. @misc{sparkx2.5, title = {Spark-X2.5 4B&1.7B: Pushing the Limits of Agentic Capabilities in On-Device Models}, author = {SparkLLM Team}, year = {2026} }