Model library / XHToken
Spark-X2.6-4B is a 4.1 B parameter causal language model from XHToken designed for on-device AI tasks.
Parameters in total
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
Spark-X2.6-4B is a 4.1 billion-parameter causal language model released by XHToken. The publisher describes it as a compact, general-purpose model aimed at on-device agents, supporting coding, multi-step planning, long-context reasoning and multilingual tasks. It retains the Spark2_5 architecture and adds post-training improvements over the previous 4B version.
In deployment the model occupies roughly 8 GB in bfloat16 format, fitting within typical GPU or NPU memory budgets. It accepts up to 1,048,576 tokens per request, enabling very long contexts. The Apache-2.0 licence permits commercial use. No expert layers are present, so inference follows a standard dense forward pass.
Specification
| Repository | XHToken/Spark-X2.6-4B |
| Publisher | XHToken |
| Published | 19 October 2026 |
| Architecture | Spark2_5ForCausalLM |
| Model type | spark2_5 |
| Parameters | 4,112,079,360 |
| Layers | 36 |
| Hidden size | 2,560 |
| Attention heads | 16, 4 key/value heads |
| Context length | 1,048,576 tokens |
| Vocabulary | 131,072 tokens |
| Weight format | bfloat16 |
| Licence | apache-2.0 |
| Base model | XHToken/Spark-X2.6-4B-Base |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 7.7 GiB | Full precision as released |
| fp8 | 3.8 GiB | 8-bit, near-lossless on most models |
| int4 | 1.9 GiB | 4-bit, smallest footprint |
The cache costs 144 KiB per token at 16-bit, so the full 1,048,576-token context of one request needs about 144.0 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses sliding-window attention with a 512-token window on some layers, so long requests cache less than this.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | 1 | 11.4 GiB | about 83,308 tokens |
| RTX 5090 32GB | 1 | 18.6 GiB | about 135,737 tokens |
| L40S 48GB | 1 | 33.0 GiB | about 240,594 tokens |
| A100 80GB | 1 | 61.8 GiB | about 450,309 tokens |
| H100 80GB | 1 | 61.8 GiB | about 450,309 tokens |
| RTX PRO 6000 96GB | 1 | 76.2 GiB | about 555,167 tokens |
| H200 141GB | 1 | 116.7 GiB | about 850,079 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve XHToken/Spark-X2.6-4B \ --max-model-len 81920 \ --gpu-memory-utilization 0.90
Questions
The weights take about 8 GiB at the released precision. That fits on one RTX 4090 24GB, which leaves roughly 11 GiB for the key-value cache.
The repository declares apache-2.0. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 1,048,576 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
Spark-X2.5-4B is a 4.1 billion-parameter causal language model from XHToken designed for general-purpose tasks with a 1 M token c…
Spark-X2.5-1.7B-Base is a 1.7 billion-parameter general-purpose language model with a native 1 M-token context window.
Spark-X2.5-4B is a 4.1-billion-parameter causal language model with a 1-million-token context window, released under Apache-2.0.
Spark-X2.5-1.7B is a 1.7 billion-parameter causal language model from XHToken with a 1 M token context window, released under Apa…
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.