The 4B model that out-agents models twice its size

What it does Spark-X2.5 is an open-weight model series from XHToken: a 4B and a 1.7B general-purpose LLM covering conversation, writing, translation, reasoning, coding, tool use, and agentic workflows across 200+ languages. The GitHub repo itself is the release hub — model cards, benchmark tables, and download pointers — while the weights live on Hugging Face, ModelScope, Ollama, and several regional mirrors, with FP8 and INT8 quantized versions available.

The interesting bit The architecture is the hook: one full-attention layer for every three sliding-window layers, which keeps long-context compute in check while natively supporting up to 1M tokens. The training bet is agentic — the 4B scores 30.4 on τ³-bench where Qwen3.5-9B manages 9.3, and 40.9 on BrowseComp against 8.3. Training ran on Huawei Ascend clusters with large-scale RL plus a post-training technique the README names MOPD but never explains.

Key highlights

• Hybrid attention (1 full + 3 sliding-window layers) delivers a native 1M-token window without the usual long-context compute tax
• Agent-first results: leads same-size and larger models on τ³-bench, MCP-Atlas, BrowseComp, SWE-Bench Pro, and Workspace Bench
• Runs in vLLM, SGLang, llama.cpp, and MLX; deployable through Ollama, LM Studio, and the PocketPal AI app on iOS and Android
• Hardware reach beyond NVIDIA — Huawei Ascend and Hygon support — with fine-tuning via LLaMA-Factory
• Integrated with agent harnesses including Codex, Claude Code, OpenClaw, and Hermes

Caveats

• This is a model-card repo, not code: no source, no training scripts — the deliverable is weights hosted elsewhere
• The benchmark table is the vendor’s own, and it is not a clean sweep — the 4B trails Qwen3.5-9B on SWE-Bench Verified (41.6 vs 53.1) and on τ²-bench
• MOPD, the headline post-training technique, gets a name and nothing more

Verdict Worth a look if you run local agents on modest hardware, especially outside the NVIDIA bubble. Skip it if you came for a repo to read — the code here is a README, and the substance is in the weights.