Is Spark-X2.6-4B state of the art?Pending benchmarks.Announced Oct 20, 2026; no independent results yet.
iFlytek SparkLLM's compact open model for on-device agents. It keeps Spark-X2.5-4B's hybrid attention design (one full-attention layer per three sliding-window layers) with a 1M-token context, and adds new post-training plus a three-level thinking budget (none, low, high). Apache 2.0, with Base, FP8, INT8, INT4 and GGUF releases.
The details so far
- Lab
- iFlytek
- Announced
- Oct 20, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- About 4.1B
- Lab
- SparkLLM team at iFlytek (Hugging Face org XHToken)
- Context
- 1,048,576 tokens
- Thinking
- Three thinking levels: none, low and high
- Lineage
- Same architecture as Spark-X2.5-4B (released Sep 1, 2026); gains come from new post-training
- Checkpoints
- Instruct and Base, plus official FP8, INT8, INT4 and GGUF builds
- License
- Apache 2.0
- Runtimes
- Transformers, SGLang, vLLM, llama.cpp, Ollama, MLX and LM Studio
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: 63.3 average on SparkLLM's reported suite (thinking set to high), against 52.1 for Spark-X2.5-4B and 49.1 for Qwen3.5-9B. The suite includes two in-house benchmarks, SWebBench and PAgentBench.
Lab claim: state of the art among similarly sized open models, matching or beating some ~9B models.
Sources
LABSparkLLM release post (DEV Community)dev.to/sparkllm/spark-x26-4b-stronger-on-device-agents-coding-and-controllable-thinking-2fh · Oct 20, 2026REPOSpark-X2.6-4B on Hugging Facehuggingface.co/XHToken/Spark-X2.6-4B · Oct 20, 2026PRESSModelScope announcement on Xx.com/ModelScope2022/status/2112781153065165291 · Oct 21, 2026Tracking continues until an independent board lists it.