iFlytek ships Spark-X2.6-4B, an open 4B model with 1M-token context
The on-device model keeps its predecessor architecture but adds selectable thinking levels and ships from day one for Ollama, llama.cpp and MLX.
iFlytek released Spark-X2.6-4B on 21 October, open-weight under Apache 2.0, with an official unveiling at its 1024 Global Developer Festival the next day. Weights are up on Hugging Face and ModelScope, with GGUF, INT4, INT8 and FP8 builds already published alongside the full-precision release.
The model has 4B parameters and a maximum context window of 1,048,576 tokens. It keeps the exact architecture of the earlier Spark-X2.5-4B, with the gains coming entirely from improved post-training. New here are three configurable thinking levels, none, low and high, which let an application trade reasoning quality against latency and inference cost without switching models. Deployment is supported out of the box through Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, Unsloth and Docker.
Why it matters: a 4B model that fits on a laptop GPU but carries a seven-figure context window is aimed squarely at coding and agent workloads you’d rather run locally than pay per token for. The thinking-level switch is a cheap way to cut latency in an agent loop during simple steps without maintaining two separate models.
The caveat: iFlytek says the model matches or beats some 9B models on several benchmarks, but the model card doesn’t publish the specific benchmark names or scores behind that claim, so treat it as a vendor statement until someone runs independent evals.