Evergreen Tools
|
← Back to AI Tools

iFlytek Spark X2.6: A 293B-A30B Cloud Model With a 4B Open-Weight Twin for Your Laptop

iFlytek released Spark X2.6 at its 1024 Global Developer Festival on 22 October 2026. The cloud model is a 293B-A30B mixture-of-experts build with a 1M-token context, trained according to the company entirely on domestic Chinese compute; alongside it, the on-device Spark-X2.6-4B ships as Apache-2.0 open weights on Hugging Face, ModelScope and Ollama

Official Links

GitHub · Hugging Face · 发布报道

Features

  • ✓ The cloud model uses a 293B total / roughly 30B active mixture-of-experts layout with context extended to 1 million tokens, aimed at long-horizon tasks and large codebases
  • ✓ iFlytek says it is the first Chinese model to reach a step change in coding and agent capability while trained entirely on domestic compute, using long-horizon dynamic parallel training adapted to domestic chips' memory bandwidth limits
  • ✓ Spark-X2.6-4B is a 4B-parameter, Apache-2.0, natively 1M-context open-weight model that keeps the X2.5-4B architecture but swaps in a new post-training recipe, built for on-device agents
  • ✓ The 4B build ships in GGUF, FP8, INT8 and INT4 quantisations, with an MLX variant for Apple Silicon and native llama.cpp architecture support in the official repo
  • ✓ iFlytek's own benchmarks put X2.6 first on SWE Pro and SWE Multilingual, second on SWE Verified, and in the top group on enterprise operations agent tasks and the MCP-Atlas tool-use test — vendor-reported, not independent

How to Use

  1. Decide cloud or device first: use the cloud API for long context and hard coding work, or pull Spark-X2.6-4B weights when offline operation and data locality matter
  2. For on-device use download XHToken/Spark-X2.6-4B from Hugging Face or ModelScope (ModelScope is faster from mainland China), or simply pull the model through Ollama
  3. Choose a quantisation for your hardware: Q4_K_M is about 2.6GB of weights and fits around 8GB of VRAM, while Apple Silicon can try the MLX 4-bit build; start with one multi-step tool-calling task
  4. Build a 60-case comparison set from your hardest work — code repair, long-document QA, multi-step tool use, 20 each — and compare accuracy and cost per run against your current model before switching

FAQ

What are the actual specs of Spark X2.6?

Per Pandaily and the launch event, the cloud model is a 293B-A30B mixture-of-experts build (293 billion total parameters, about 30 billion active per token) with context up to 1 million tokens, trained entirely on domestic compute; iFlytek calls it the first Chinese model to reach a step change in coding and agent capability on domestic silicon. The family also includes the open-weight on-device build published under the XHToken organisation on Hugging Face. Code at https://github.com/XHToken/Spark-X2.6 and weights at https://huggingface.co/XHToken/Spark-X2.6-4B

What can the open 4B version actually do?

The model card describes it as a compact 4B open-weights model for resource-conscious on-device agentic deployment, Apache 2.0, with native context of 1,048,576 tokens. It keeps the Spark-X2.5-4B architecture and adds an improved post-training recipe, and iFlytek claims substantial gains in software engineering, multi-step planning, complex tool use, math, multilingual understanding and ultra-long context, saying it rivals or surpasses larger 9B models. Those are vendor benchmarks — pressure-test them yourself

How do I run it locally and how much VRAM does it need?

Official GGUF quantisations are published, with Q4_K_M around 2.6GB and Q5_K_M around 3.16GB of weights; the community also maintains quant tables (bartowski) and an MLX 4-bit build (abenzerps). The official repo supports llama.cpp, which has had native Spark2_5ForCausalLM support since 6 September 2026, plus MLX, running on Apple Silicon, Linux CPU and NVIDIA CUDA. Actual VRAM depends heavily on context length — a full million-token window will eat memory, so start short

What does it cost?

Pandaily reports limited-time pricing of 0.9 yuan per million input tokens and 0.08 yuan per million cached input tokens, with the output rate cut off in the report — check the official iFlytek open platform page for the full table. The 4B on-device build is Apache-2.0 open weights, so local inference costs nothing beyond hardware and power

What workloads suit it?

Cloud X2.6 fits long-context codebase Q&A, multi-step agent orchestration and enterprise flows that need tool calling; the on-device 4B fits offline, privacy-sensitive or flaky-network settings such as a local document assistant, an edge agent or intranet code completion. The usual rule applies: identify your single most painful task, run 20 real samples against it, and ignore leaderboards until your own numbers agree