Spark-X2.6
TokenSpark's 1.7B and 4B on-device agent models with switchable thinking and a 1M-token position limit. Apache 2.0.
In short. Spark-X2.6 is a pair of small language models, 1.7B and 4B parameters, built to run on a laptop, a phone or an edge board while still using tools and writing code. You can turn its step-by-step thinking off, low or high to trade quality for speed. It is free to use, including commercially, under Apache 2.0.
Summary
Spark-X2.6 is the second open-weight family of compact language models from TokenSpark, published on Hugging Face on 19 October 2026 under Apache 2.0. It comes in 1.7B and 4B sizes, each with a base checkpoint, FP8, INT8 and INT4 builds and GGUF files. By its configuration, the 4B model has 36 layers, in which three of every four use sliding-window attention over 512 tokens and the fourth uses full attention, and its maximum position setting is 1,048,576 tokens. The card's SGLang example uses that as the context length and warns that it needs enough device memory. TokenSpark says Spark-X2.6 keeps the architecture of the August Spark-X2.5 and gains from better post-training, mainly in coding, agentic tool use, long context, multilingual use and reasoning.
Thinking is switchable. The chat template defaults to high effort, you can set low, or you can turn thinking off. The card reports that moving from high to low costs 1.7 points of average score and cuts output length by 22.0% for the 4B model, and costs 4.0 points and cuts length by 37.3% for the 1.7B model. Its benchmark charts are images, so we quote none of the scores. TokenSpark claims the best results among similar-sized open models, matching or beating some 9B models on several benchmarks, and efficiency equal to or better than Qwen3.5-4B and Gemma-E4B. The card documents Transformers with remote code, vLLM and SGLang on NVIDIA GPUs and Ascend NPUs, Spark-MLX-LLM for Apple silicon, Linux CPU and CUDA, GGUF files for llama.cpp, Ollama and LM Studio, and a community WebGPU demo that runs the INT4 build in a browser.
The 4B model weighs 8.24 GB in BF16 and 3.41 GB in INT4, and the 1.7B model weighs 3.43 GB and 1.71 GB. In the 30 days to 11 November 2026, Hugging Face counted about 660,000 downloads of the 4B GGUF repository and about 27,000 of the 4B Transformers repository. The card recommends Llama-Factory for fine-tuning. We found no independent evaluation and little public information about the company beyond its Hugging Face profile and website, so test it on your own tasks.
Good for
- On-device assistants and agents on laptops, phones and edge boards, since the 4B INT4 build is 3.4 GB
- Fast local chat through Ollama, LM Studio or llama.cpp with the GGUF files
- Apps that want a thinking switch, so easy requests can skip the long reasoning
- Fine-tuning a small Apache 2.0 model with Llama-Factory
Watch out for
- The benchmark results are charts from TokenSpark, so check the model on your own tasks before relying on its claims
- A million-token context needs a lot of memory, and the card says to reduce the context length when necessary
- The card's example loads the model with remote code and pins Transformers 4.57.1
- Little is publicly known about the company, so check the track record and support you need
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Spark-X2.6-1.7B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Spark-X2.6-1.7B BF16 | 1.71B | 3.93 GB | 1.97 GB | 1.18 GB | Laptop or phone |
| Spark-X2.6-1.7B INT4 INT4 | 1.71B | 3.93 GB | 1.97 GB | 1.88 GB | Laptop or phone |
| Spark-X2.6-4B BF16 | 4.11B | 9.45 GB | 4.73 GB | 2.84 GB | Laptop or phone |
| Spark-X2.6-4B INT4 INT4 | 4.11B | 9.45 GB | 4.73 GB | 3.75 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is Spark-X2.6?
- Spark-X2.6 is a pair of small language models, 1.7B and 4B parameters, built to run on a laptop, a phone or an edge board while still using tools and writing code. You can turn its step-by-step thinking off, low or high to trade quality for speed. It is free to use, including commercially, under Apache 2.0.
- Can I use Spark-X2.6 commercially?
- Yes. Spark-X2.6 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Spark-X2.6 open source?
- Spark-X2.6 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Spark-X2.6 on my own computer?
- Yes. The smallest variant, Spark-X2.6-1.7B, needs about 1.18 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Spark-X2.6 released?
- Spark-X2.6 was released by TokenSpark (SparkLLM) on October 19, 2026.
- How long a context window does Spark-X2.6 have?
- Spark-X2.6 supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Spark-X2.6-4B model card (Hugging Face) huggingface.co
- Spark-X2.6-4B configuration file (Hugging Face) huggingface.co
- Spark-X2.6-4B Apache 2.0 license file (Hugging Face) huggingface.co
- Spark-X2.6-1.7B model page (Hugging Face) huggingface.co
- XHToken organization profile (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| LFM3 Liquid AI | 1.73B – 2.17B | 256K | LFM Open 1.0 (Conditional) | 1.19 GB Laptop or phone | |
| Nemotron 3 Nano 4B NVIDIA | 3.97B | 256K | NVIDIA Nemotron (Permissive) | 2.74 GB Laptop or phone |