API ManyНовые модели: Qwen 3.8 Max, Claude Haiku 5.5, Claude Opus 5.5, Claude Sonnet 5.5
SparkLLMGuide

Spark-X2.6-4B: the 4B open-weight Spark model, explained

Spark-X2.6-4B is a 4B Apache-2.0 model with a 1M-token context and three thinking levels, built for on-device agents. Specs, published benchmarks against Qwen3.5 and Gemma 4, and where to run it.

30 окт. 2026 г. · API Many

Spark-X2.6-4B is a compact, general-purpose language model for on-device use. It was uploaded to Hugging Face on October 19, 2026, and its publisher announced the global launch of the Spark-X2.6 series on Hugging Face, ModelScope and Ollama on October 21. It keeps exactly the same architecture as Spark-X2.5-4B, released on September 1, 2026; the gains come from better post-training, with the biggest claimed improvements in coding, agents, long context and multilingual work.

Sources:Hugging Face: XHToken/Spark-X2.6-4B model card · Hugging Face: XHToken (SparkLLM) organization page

Who makes it

The weights are published by the Hugging Face organization XHToken, whose display name is SparkLLM; it describes itself as the maker of the Spark model family and the AStudio and Loomy agent products. iFlytek’s Spark (讯飞星火) API page lists the same model, Spark-X2.6-4B, as a 4B on-device dense model, so it belongs to iFlytek’s Spark family. The Hugging Face organization is not marked verified and does not name a company.

Sources:Hugging Face: XHToken (SparkLLM) organization page · iFlytek: Spark API models and prices (Chinese)

Do not confuse it with Spark-X2.6 without a size: on iFlytek’s API that name is a hosted 293B-A30B mixture-of-experts model with a 1M-token context, priced at ¥0.9 input and ¥3.6 output per 1M tokens during a limited-time 40%-off offer. This page covers only the 4B open-weight model.

Source:iFlytek: Spark API models and prices (Chinese)

Spark-X2.6-4B at a glance

Spark-X2.6-4BSource
On Hugging FaceOctober 19, 2026 (series launch announced October 21)Organization page
Parameters4.11B (BF16), denseModel card
Architecture36 layers: 27 sliding-window (512-token window) and 9 full-attention layers, the same as Spark-X2.5-4Bconfig.json
Context window1,048,576 tokensconfig.json
ThinkingThree levels: none, low, high. On by default; turn off with enable_thinking=falseModel card
LanguagesOver 200, per the publisher; the card is tagged English and ChineseOrganization page
LicenseApache 2.0Model card
VariantsBase, FP8, INT8, INT4 and GGUF checkpoints, plus a smaller Spark-X2.6-1.7BHugging Face
Hosted APINone from the major routers; listed as free for a limited time on iFlytek’s Xingchen MaaSiFlytek
On API ManyNoSee the note above

Published benchmarks

BenchmarkSpark-X2.6-4BSpark-X2.5-4BQwen3.5-4BQwen3.5-9BGemma 4 E4B
Average (publisher’s chart)63.352.143.249.129.7
BFCL-V4 (function calling)71.865.150.3*66.1*47.0*
τ²-bench86.975.179.9*79.1*42.2*
MCP-Atlas62.454.640.8*47.4*15.0*
SWE-bench Verified62.841.638.8*53.1*14.0*
SWE-bench Pro54.344.429.4*33.8*4.0*
AIME 202691.290.783.088.242.5*
AA-LCR v1.1 (long context)67.364.763.0*70.0*32.0*
Graphwalks BFS 1M (F1)34.22.515.825.3–
GPQA Diamond77.173.776.2*81.7*58.6*
Humanity’s Last Exam (no tools)17.912.38.614.33.9
IFEval93.693.089.8*91.5*45.3
Numbers are the publisher’s own, from the model card’s benchmark chart. * marks results the publisher copied from other model cards or papers; – means no score. Spark-X2.6-4B was evaluated in thinking mode at the high level. The chart also includes Gemma 4 12B and two in-house benchmarks (SWebBench, PAgentBench) that are not yet public.

Source:Spark-X2.6-4B model card: full benchmark chart

The publisher’s claim is that the 4B model matches or beats 9B-class models on several tasks. Its own chart supports that for agent and tool-use benchmarks, coding and very long context, while Qwen3.5-9B stays ahead on GPQA Diamond and AA-LCR. These are the publisher’s own measurements, not independent results.

Sources:Hugging Face: XHToken/Spark-X2.6-4B model card · Spark-X2.6-4B model card: full benchmark chart

How to run it

  • Transformers: version 4.57.1 with trust_remote_code=True. Recommended sampling is temperature 1.0, top_p 0.95 and no top_k limit.
  • Servers: SGLang and vLLM (NVIDIA GPUs and Ascend NPUs) both expose an OpenAI-compatible API on your own machine. The card’s example sets the full 1,048,576-token context, which needs a lot of memory; reduce it if needed.
  • Local apps: llama.cpp build b10828 or later, Ollama 0.34.1 or later, LM Studio 0.4.0 or later, Unsloth Studio, and Apple MLX through the publisher’s Spark-MLX-LLM.
  • Fine-tuning: the publisher recommends LLaMA-Factory.

Sources:Hugging Face: XHToken/Spark-X2.6-4B model card · Ollama: SparkLLM/Spark-X2.6-4B

ollama run SparkLLM/Spark-X2.6-4B

What to use on API Many instead

If you want a small, low-cost model as a hosted API rather than weights to run yourself, API Many offers Qwen 3.8 Flash from $0.114 вход · $0.386 выход, GPT-6 Luna from $0.10 вход · $0.50 выход and Gemini 3.5 Flash Lite from $0.30 вход · $2.499 выход, all through one OpenAI-compatible endpoint.

Related

FAQ

Who released Spark-X2.6-4B?

It is published on Hugging Face by the XHToken organization (display name SparkLLM), and iFlytek’s Spark API page lists the same model as part of its Spark family. The Hugging Face organization is not verified and does not name a company.

Is Spark-X2.6-4B open source?

Yes. The weights are released under the Apache 2.0 license, along with a base model and FP8, INT8, INT4 and GGUF versions.

How long is its context window?

1,048,576 tokens (about 1M), according to its configuration and the publisher’s serving examples.

Does it support thinking mode?

Yes. It has three thinking levels (none, low and high) and thinks by default. Set enable_thinking to false to turn thinking off.

Is there a hosted API for Spark-X2.6-4B?

Not from Hugging Face Inference Providers, OpenRouter or Vercel AI Gateway as of October 30, 2026. iFlytek lists it on its Xingchen MaaS platform as free for a limited time; otherwise you run it yourself with vLLM, SGLang, llama.cpp, Ollama or LM Studio.

Is it the same as Spark-X2.6?

No. On iFlytek’s API, Spark-X2.6 is a much larger hosted 293B-A30B mixture-of-experts model. Spark-X2.6-4B is the 4B open-weight model for devices.

Can I use Spark-X2.6-4B on API Many?

Not currently. For a small hosted model on API Many, try Qwen 3.8 Flash, GPT-6 Luna or Gemini 3.5 Flash Lite.

Sources

  1. Hugging Face: XHToken/Spark-X2.6-4B model card — Description, thinking levels, license, run and fine-tuning instructions.
  2. Spark-X2.6-4B model card: full benchmark chart — All benchmark numbers and evaluation notes.
  3. Spark-X2.6-4B config.json — Layer layout, sliding window and maximum context length.
  4. Hugging Face: XHToken (SparkLLM) organization page — Publisher description, launch dates and language coverage.
  5. Hugging Face: XHToken/Spark-X2.6-1.7B — The Spark-X2.6-1.7B sibling.
  6. iFlytek: Spark API models and prices (Chinese) — Spark-X2.6-4B MaaS listing and the separate hosted Spark-X2.6 (293B-A30B) model and price.
  7. Ollama: SparkLLM/Spark-X2.6-4B — Ollama model name.
  8. Hugging Face: Inference Providers — Hugging Face Inference Providers (none serve this model as of Oct 30, 2026).
Попробуйте с одним API-ключом

Все модели из статьи доступны в API Many. Оплата по факту от $10.

Смотреть модели →

Ещё в блоге