Spark-X2.6 4B
Spark-X2.6 4B needs 3.3 GB of VRAM at Q4_K_M with 8k context, computed from its config.json as checked on 3 November 2026. The smallest card that holds it has 8 GB, and the fastest desktop card that runs it, the RTX 5090, generates at most 528 tokens per second.
- Parameters
- 4.1B
- Context
- 1,048,576 tokens
- Licence
- Apache 2.0
- Released
- October 2026
- Config checked
- 3 November 2026
Open Spark-X2.6 4B in the LLM checker
Which card size runs it
3.3 GB against each memory size. A model may use 95% of a card, and only what is left over, or missing, is coloured; tight means above 85%, with little room for a longer conversation.
Other quantisations, context lengths or a Q8 KV cache: the VRAM calculator.
Computed 3 November 2026Which GPU to pick
The smallest card here that loads it at Q4_K_M with 8k of context is the RTX 3060 (12 GB): 3.3 GB needed, an estimated 106 tokens per second. That is only the smallest card with its own page here: the model itself fits any 8 GB card. The same card still has room at 32,768 tokens of context (4.2 GB), so there is no reason to buy above it for this model.
The same card also holds it at Q8_0, in 5.4 GB. Q8 is near-lossless, so take the higher precision when it costs nothing. On Apple silicon, a Mac mini M4 Pro holds it with 32,768 tokens of context at an estimated 80 tokens per second; prompt processing is slower than on Nvidia, which long documents make obvious. Among unified-memory desktops, the Ryzen AI Max+ 395 holds it at 32,768 tokens in 4.2 GB, at an estimated 75 tokens per second.
Speed on each card
| GPU | Fits | Est. tokens/s | Feels like | Memory | Bandwidth |
|---|---|---|---|---|---|
| H100 SXM | Yes | ≤ 986 | faster than you can read | 80 GB | 3,350 GB/s |
| A100 | Yes | ≤ 600 | faster than you can read | 80 GB | 2,039 GB/s |
| RTX 5090 | Yes | ≤ 528 | faster than you can read | 32 GB | 1,792 GB/s |
| RTX PRO 6000 Blackwell | Yes | ≤ 528 | faster than you can read | 96 GB | 1,792 GB/s |
| RTX PRO 5000 Blackwell | Yes | ≤ 396 | faster than you can read | 48 GB | 1,344 GB/s |
| RTX 4090 | Yes | ≤ 297 | faster than you can read | 24 GB | 1,008 GB/s |
| RTX 5080 | Yes | ≤ 283 | faster than you can read | 16 GB | 960 GB/s |
| RX 7900 XTX | Yes | ≤ 283 | faster than you can read | 24 GB | 960 GB/s |
| RTX 6000 Ada | Yes | ≤ 283 | faster than you can read | 48 GB | 960 GB/s |
| RTX 3090 | Yes | ≤ 276 | faster than you can read | 24 GB | 936 GB/s |
| RTX 5070 Ti | Yes | ≤ 264 | faster than you can read | 16 GB | 896 GB/s |
| RTX 5090 Laptop GPU | Yes | ≤ 264 | faster than you can read | 24 GB | 896 GB/s |
| RTX PRO 4500 Blackwell | Yes | ≤ 264 | faster than you can read | 32 GB | 896 GB/s |
| L40S | Yes | ≤ 254 | faster than you can read | 48 GB | 864 GB/s |
| RX 7900 XT | Yes | ≤ 235 | faster than you can read | 20 GB | 800 GB/s |
| RTX A6000 | Yes | ≤ 226 | faster than you can read | 48 GB | 768 GB/s |
| RTX 4080 SUPER | Yes | ≤ 217 | faster than you can read | 16 GB | 736 GB/s |
| RTX 5070 | Yes | ≤ 198 | faster than you can read | 12 GB | 672 GB/s |
| RTX PRO 4000 Blackwell | Yes | ≤ 198 | faster than you can read | 24 GB | 672 GB/s |
| RX 9070 XT | Yes | ≤ 188 | faster than you can read | 16 GB | 640 GB/s |
| Radeon AI PRO R9700 | Yes | ≤ 188 | faster than you can read | 32 GB | 640 GB/s |
| Arc Pro B70 | Yes | ≤ 179 | faster than you can read | 32 GB | 608 GB/s |
| RTX 4070 | Yes | ≤ 148 | faster than you can read | 12 GB | 504 GB/s |
| Arc B580 | Yes | ≤ 134 | faster than you can read | 12 GB | 456 GB/s |
| Arc Pro B60 | Yes | ≤ 134 | faster than you can read | 24 GB | 456 GB/s |
| RTX 5060 Ti | Yes | ≤ 132 | faster than you can read | 16 GB | 448 GB/s |
| RTX PRO 4000 Blackwell SFF | Yes | ≤ 127 | faster than you can read | 24 GB | 432 GB/s |
| RTX 3060 | Yes | ≤ 106 | faster than you can read | 12 GB | 360 GB/s |
| RX 9060 XT | Yes | 94 | faster than you can read | 16 GB | 320 GB/s |
| RTX Spark | Yes | 88 | faster than you can read | 128 GB (102 for the GPU) | 300 GB/s |
| RTX 4060 Ti | Yes | 85 | faster than you can read | 16 GB | 288 GB/s |
| DGX Spark | Yes | 80 | faster than you can read | 128 GB (126 for the GPU) | 273 GB/s |
| Ryzen AI Max+ 395 | Yes | 75 | faster than you can read | 128 GB (96 for the GPU) | 256 GB/s |
Single-stream decode, from memory bandwidth at 70% efficiency. Prompt processing and batching not included. These are ceilings, not measurements: Spark-X2.6 4B reads only 2.4 GB of weights per token at Q4, so little that the per-token costs the formula leaves out decide much of the real speed. The 28 figures marked ≤ are where a real runtime falls furthest below the number shown. See the speed estimator for other quantisations and Apple hardware, or every card compared if you are choosing hardware rather than a model.
Computed 3 November 2026Memory by quantisation and context
| Quant | 4,096 ctx | 8,192 ctx | 32,768 ctx |
|---|---|---|---|
| Q8_0 | 5.2 GB · fits 8 GB | 5.4 GB · fits 8 GB | 6.3 GB · fits 8 GB |
| Q5_K_M | 3.7 GB · fits 8 GB | 3.9 GB · fits 8 GB | 4.8 GB · fits 8 GB |
| Q4_K_M | 3.2 GB · fits 8 GB | 3.3 GB · fits 8 GB | 4.2 GB · fits 8 GB |
Weights at this quant: 2.4 GB at Q4. Every extra 1,000 tokens of context adds about 0.037 GB of KV cache at FP16, on top of a fixed 0.057 GB that the sliding-window layers hold once their window is full. Try other settings in the VRAM calculator.
About the model
Spark-X2.6 4B has 4.1B parameters. It has 36 layers: 27 use sliding-window attention over the last 512 tokens and 9 are full attention, all with 4 KV heads of dimension 256. The SparkLLM team's on-device model from October 2026, published under XHToken on Hugging Face: a 4B text model whose release notes put agent work first, meaning tool use, multi-step planning and software engineering. It keeps Spark-X2.5 4B's architecture, with three of every four layers attending only to the last 512 tokens, and its config allows a context of 1,048,576 tokens.
Text only 1,048,576-token window Apache 2.0 October 2026
Reasoning. Three levels, none, low and high. On by default, and switched off per request in the chat template.
A runtime without a sliding-window cache stores every layer at full length and needs far more than this at long context.
Native context window: 1,048,576 tokens. Licence: Apache 2.0, a permissive licence. Weights published in October 2026.
Good at
- Agents on small hardware: tool use and multi-step planning are what the card says this release improved.
- Very long inputs on a small card: only 9 of its 36 layers keep a cache that grows, and its window reaches 1,048,576 tokens.
- Runtime support named on the card from day one: llama.cpp, Ollama, LM Studio and MLX.
- Apache 2.0.
Watch out for
- The card's server examples open the full 1,048,576-token context and warn that it "requires sufficient device memory". Set a shorter context unless you need it.
- It needs a recent runtime: llama.cpp b10828 or later, Ollama 0.34.1 or later, or LM Studio 0.4.0 or later. Transformers is pinned to version 4.57.1, with remote code.
- The card lists English and Chinese as its languages.
- It is still a 4B model. The long window lets it read a lot; it does not make it reason like a 27B.
Similar models
The models that need about the same memory as Spark-X2.6 4B at Q4_K_M and 8k, which makes them the real alternatives on whatever card you have. The context column is what each extra 1,000 tokens costs, and it is where models of the same size differ most.
| Model | Size | Needs | Per 1k context | Window | Released |
|---|---|---|---|---|---|
| Spark-X2.6 4B | 4.1B | 3.3 GB | 0.037 GB | 1024k | 2026-10 |
| Llama 3.2 3B | 3.2B | 3.4 GB | 0.115 GB | 128k | 2024-09 |
| Qwen3.5 4B | 4.7B | 3.7 GB | 0.033 GB | 256k | 2026-02 |
| Llama 3.1 8B | 8B | 6.4 GB | 0.131 GB | 128k | 2024-07 |
| Qwen3 8B | 8.2B | 6.7 GB | 0.147 GB | 40k | 2025-04 |
See every model compared, which models fit on 8 to 96 GB, or the cheapest hour of every cloud card.
Questions
- Can I run Spark-X2.6 4B on a 24 GB card?
- Yes, at Q4_K_M and 8k context it needs about 3.3 GB, which fits a 24 GB card with room.
- How much VRAM does Spark-X2.6 4B need at Q8?
- About 5.4 GB at 8k context, or 6.3 GB at 32k. Q8 is near-lossless; use it when it fits.
- What is the context length of Spark-X2.6 4B?
- 1,048,576 tokens natively. Holding all of it at Q4_K_M takes about 41.7 GB, of which 38.7 GB is context.
- Why does Spark-X2.6 4B need so little memory for long context?
- Because most of its layers do not keep a cache that grows. It has 36 layers: 27 use sliding-window attention over the last 512 tokens and 9 are full attention, all with 4 KV heads of dimension 256. At 32k tokens the context costs about 1.3 GB; if all 36 layers kept a conventional full-length cache, the same conversation would cost about 4.8 GB. That saving depends on the runtime implementing the sliding-window cache. llama.cpp and vLLM do; a build that does not will use far more.
- How fast is Spark-X2.6 4B on an RTX 4090?
- At most about 297 tokens per second at Q4, single stream, which is faster than you can read. That is a ceiling worked out from the card's memory bandwidth, not a measurement, and at a figure this high real runtimes land well below it, because the costs the formula leaves out take a growing share of each token.
- Architecture values from config.json in XHToken/Spark-X2.6-4B on Hugging Face, and its model card, checked 3 November 2026. Every figure on this page is computed from those values; check them before buying hardware for this model.
- Estimates assume a single conversation on a card that is otherwise free. A desktop on the same GPU takes 0.5 to 2 GB.