Spark-X2.6-4B is a new 4B open model from the SparkLLM Team, and it scores well above its size on agent and coding tests. It has a 1M-token context window, three thinking levels and an Apache 2.0 licence. It came out a little over a month after Spark-X2.5-4B and uses the same architecture, so the gains come from better post-training. In the video below I test it on real jobs. Here is what it does, how it compares and how to run it.
Key takeaways
- 4B parameters, Apache 2.0, English and Chinese.
- Average benchmark score 63.3 vs 52.1 for Spark-X2.5-4B.
- 1M context, if your device has enough memory.
- Runs in Ollama, LM Studio, llama.cpp, SGLang, vLLM and MLX.
What is Spark-X2.6-4B?
Spark-X2.6-4B is a small language model from the SparkLLM Team, published on Hugging Face under the XHToken organisation. The model card says it keeps the same architecture as Spark-X2.5-4B, with the improvement coming from post-training. There is also a base checkpoint, Spark-X2.6-4B-Base, if you want to fine-tune it.
| Spark-X2.6-4B at a glance | |
|---|---|
| Size | 4B parameters, BF16 |
| Layers | 36, with 3 sliding-window layers to every 1 full attention layer |
| Context | Up to 1,048,576 tokens |
| Thinking | none, low or high (on by default) |
| Languages | English and Chinese |
| Licence | Apache 2.0 |
Benchmarks vs Qwen3.5 and Gemma4
All scores below are from the model card, in thinking mode, with Spark-X2.6 on high. Some rival scores were taken from those models’ own cards.
| Benchmark | Spark-X2.6-4B | Spark-X2.5-4B | Qwen3.5-9B | Qwen3.5-4B |
|---|---|---|---|---|
| Average | 63.3 | 52.1 | 49.1 | 43.2 |
| BFCL-V4 (tool calls) | 71.8 | 65.1 | 66.1 | 50.3 |
| tau2 (agents) | 86.9 | 75.1 | 79.1 | 79.9 |
| MCP-Atlas | 62.4 | 54.6 | 47.4 | 40.8 |
| BrowseComp | 53.1 | 40.9 | 8.3 | 14.3 |
| SWE-bench Verified | 62.8 | 41.6 | 53.1 | 38.8 |
| SWE-bench Pro | 54.3 | 44.4 | 33.8 | 29.4 |
| AIME 2026 | 91.2 | 90.7 | 88.2 | 83.0 |
| GPQA Diamond | 77.1 | 73.7 | 81.7 | 76.2 |
| IFEval | 93.6 | 93.0 | 91.5 | 89.8 |
On the average, Gemma4-12B scores 47.0 and Gemma4-E4B 29.7. The standout is agent work: tool calls, browsing and SWE-bench are well ahead of the 9B Qwen model. It isn’t best at everything, though. Qwen3.5-9B still wins on GPQA Diamond and on some long-context tests.
Want small AI models running cheap jobs for your business? Inside the AI Profit Boardroom you get the Agent OS — Hermes, OpenClaw, OpenCode, Claude and Codex in one place — plus step-by-step Hermes training and four live coaching calls a week.
Want a plan for your business first? Claim your free strategy session.
Join the AI Profit Boardroom →Claim Your Free SEO Strategy Session →
The 1M context window
The model supports 1,048,576 tokens of context, and the server examples use that full length. The card adds a fair warning: you need enough device memory, so lower the context length if you run out. On long-context tests it does well on Graphwalks 1M (34.2 vs 2.5 for X2.5) but trails Qwen3.5-9B on MRCR v2 at 512K–1M (26.0 vs 34.3).
Want small, cheap AI models doing real work for you? Claim your free strategy session →
Thinking levels
Spark-X2.6-4B has three thinking levels: none, low and high. Thinking is on by default. To turn it off, pass enable_thinking: false in the chat template settings. The recommended sampling is temperature 1.0, top-p 0.95 and top-k -1. Use high for hard maths or coding, and none for quick replies.
How to run Spark-X2.6-4B
The model card lists all of these:
- Ollama (v0.34.1+):
ollama run SparkLLM/Spark-X2.6-4B - LM Studio (0.4.0+):
lms chat spark-x2.6 - llama.cpp (b10828+):
llama serve -m /path/to/Spark-X2.6-4B-GGUF - SGLang: the lmsysorg/sglang Docker image with
--reasoning-parser qwen3 --tool-call-parser spark25 - vLLM: with the Spark plugin from GitHub (XHToken/Spark-plugin)
- Transformers:
pip install transformers==4.57.1andtrust_remote_code=True - MLX on Mac: via XHToken/Spark-MLX-LLM
- Fine-tuning: via XHToken/LlamaFactory
It targets NVIDIA GPUs, Ascend NPUs, Apple silicon and Linux CPUs. The card doesn’t give a minimum hardware spec, but a 4B model at 4-bit is small enough for most modern laptops at normal context lengths.
What I’d use it for
In my video I look at the kind of everyday jobs a small model is good at:
- Onboarding new team members: load your docs and let it answer questions.
- Summarising long documents: the 1M context means you can paste a lot in one go.
- Cheap agent tasks: its tool-calling scores make it a good worker model behind a bigger planner.
- Private work: it runs fully on your own machine.
Be realistic about limits. It is still a 4B model, so for hard reasoning or big code changes, a larger model will do better. Two of the in-house benchmarks, SWebBench and PAgentBench, are not public yet; the team says they will be open-sourced soon.
Want a shorter take? I also cover Spark-X2.6-4B on the AI Profit Boardroom blog.
Spark-X2.6-4B vs Spark-X2.5-4B
The architecture is the same, so everything you learned with Spark-X2.5-4B carries over. The difference is post-training, and the jump is large for a release that came only a little over a month later:
| Benchmark | X2.6 | X2.5 | Change |
|---|---|---|---|
| Average | 63.3 | 52.1 | +11.2 |
| SWE-bench Verified | 62.8 | 41.6 | +21.2 |
| BrowseComp | 53.1 | 40.9 | +12.2 |
| tau2 | 86.9 | 75.1 | +11.8 |
| MCP-Atlas | 62.4 | 54.6 | +7.8 |
| Graphwalks 1M | 34.2 | 2.5 | +31.7 |
If you already run X2.5, upgrading is a simple model swap. Your prompts and tools should keep working.
Who is Spark-X2.6-4B for?
It is a good fit if you want an AI model that runs on your own laptop, costs nothing per message and keeps your data private. Small business owners can use it to answer team questions from a handbook, sum up long reports or sort incoming messages. Developers can use it as a cheap worker model inside an agent, where a larger model plans and Spark carries out the tool calls.
It is not the right pick if you need the very best answers on hard reasoning or large code changes. For that, a bigger model will still do better. The smart move is to try Spark on your simplest daily jobs first, see how far it gets, then decide which tasks to hand to it for good.
How to get good results from a small model
- Give it one clear job at a time instead of a long list of tasks.
- Use high thinking for maths, code and multi-step tool use, and none for quick chat.
- Pair it with a bigger planner. Let a large model plan and Spark do the cheap tool calls.
- Watch your memory at long contexts. Start at 32K–128K and only go higher when you need it.
- Test it on your own data, because benchmarks are only a guide.
Used like this, a 4B model can handle a surprising amount of daily work for close to zero cost.
My verdict on Spark-X2.6-4B
For a 4B model, the agent and coding scores are very strong, and it is easy to run in Ollama or LM Studio. If you want a small, private, free model for summaries, tool calls and support tasks, Spark-X2.6-4B is one of the best small models to try right now.
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency and the AI Profit Boardroom. I test new AI agents almost every day for 432K+ YouTube subscribers and run Hermes, OpenClaw, OpenCode, Claude and Codex side by side in my own Agent OS. Claim your free strategy session or join the AI Profit Boardroom.
FAQ: spark-x2.6-4b
Who made Spark-X2.6-4B?
The SparkLLM Team, which publishes its models on Hugging Face under the XHToken organisation.
What is Spark-X2.6-4B’s context window?
Up to 1,048,576 tokens (1M), as long as your device has enough memory.
How do I run Spark-X2.6-4B in Ollama?
Use Ollama v0.34.1 or later and run: ollama run SparkLLM/Spark-X2.6-4B.
Is Spark-X2.6-4B better than Qwen3.5-9B?
On the model card’s average it scores 63.3 vs 49.1, and it leads on tool calls, browsing and SWE-bench. Qwen3.5-9B still wins on GPQA Diamond and some long-context tests.
Can I turn off thinking?
Yes. It has none, low and high levels. Pass enable_thinking false in the chat template settings to turn thinking off.
Ready to put Hermes to work in your business? Join the AI Profit Boardroom for the Agent OS, setup training and daily updates.
Or claim your free strategy session and we’ll map out your setup.
Starting from zero? Get the free AI course and community.
Join the AI Profit Boardroom →Claim Your Free SEO Strategy Session →Get the FREE AI Course + Community →
Related reading
Last updated October 2026. This page is a living guide to spark-x2.6-4b — the facts here move fast and it is updated as they do.
