[whatmodel.sh](../../README.md) / Spark-X2.5 4B
# Spark-X2.5 4B
iFLYTEK (SparkLLM) · released 2026-08 · 4.11B params · apache-2.0
**score 64.7** ±0.7 · #147 of 732 · from 34 benchmark results across 2 sources · † no independently run results verified
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Spark-X2.5 4B among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▁▁▃▆█▄●▅▄▄▂▁ #132 of 313 35% 10
coding ▁▁▁▃▅▇█●▅▄▃▁ #143 of 469 23% 4
chat & writing ▁▁▁▃▆▄▆█▆●▇▃ #102 of 402* 1.4% 1
math ▁▁▃▅█▆▄●▄▂▁▁ #93 of 431 23% 3
web search ▅·▅▅█·▃▃▃·●▃ #2 of 15 4.8% 2
agent tasks · #132 of 313 · 35% of its score
Beats [o3](o3.md), [Gemini 2.5 Pro](gemini-2-5-pro.md), [Gemini 3.1 Flash-Lite](gemini-3-1-flash-lite.md) here, though they rank higher overall. Trails [SmolLM3 3B](smollm3-3b.md), [Jamba 1.5 Large](jamba-1-5-large.md), [Devstral 2 (2512)](devstral-2.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
SWE-bench Pro (split unspecified)… 44.4% ▂·▁▁▁▁▂▃●▇▂▁ beats 23/58 very hard · avg 9.8% 9.0% sharp · at its level · several runs
SWE-bench Verified (maker) 41.6% ▁▁▁▁▁▁▂●█▆▁▁ beats 29/138 very hard · avg 14% 5.7% sharp · at its level · small run · seve…
SWE-bench Multilingual (maker) 53.3% ▁··▁▃▁●██▂▂▁ beats 13/65 hard · avg 18% 4.5% at its level · several runs
BrowseComp (maker) 40.9% ▅···█·▃···▃● beats 7/7 hard · avg 21% 3.2% at its level · several runs
τ²-bench aggregate (maker) 75.1% ▁▁▁·▁▁▃▃●▄█▁ beats 17/39 medium · avg 54% 3.1% several runs
BFCL V4 (December 2025 archive) (… 65.1% ▁▁▁▂▆▆▅▆█▂●▃ beats 99/114 medium · avg 40% 2.8% at its level · several runs
MCP-Atlas (public set) (maker) 54.6% ▂·▁▂▂·▁●█▆▆▂ beats 9/39 hard · avg 21% 2.4% at its level
MCPMark (maker) 14.2% ▃·····●··▃█▃ beats 3/8 very hard · avg 5.5% 1.9% typical
Humanity's Last Exam with tools (… 21.8% ▁·▁▂●▁▄▄▇█▄▅ beats 8/62 very hard · avg 11% 1.7% at its level
τ³-bench (maker) 30.4% ▃·▃·█·▃··▃·● beats 7/7 very hard · avg 13% 1.1% at its level
coding · #143 of 469 · 23% of its score
Beats [Qwen3.5 Plus](qwen3-5-plus.md), [Qwen3-Coder-Next (80B-A3B)](qwen3-coder-next.md), [Gemini 2.5 Pro](gemini-2-5-pro.md) here, though they rank higher overall. Trails [Llama 3.1 Nemotron Ultra 253B v1](llama-3-1-nemotron-ultra-253b.md), [Devstral Small 2 24B](devstral-small-2.md), [Gemma 4 12B](gemma-4-12b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
SWE-bench Pro (split unspecified)… 44.4% ▂·▁▁▁▁▂▃●▇▂▁ beats 23/58 very hard · avg 9.8% 9.0% sharp · at its level · several runs
SWE-bench Verified (maker) 41.6% ▁▁▁▁▁▁▂●█▆▁▁ beats 29/138 very hard · avg 14% 5.7% sharp · at its level · small run · seve…
SWE-bench Multilingual (maker) 53.3% ▁··▁▃▁●██▂▂▁ beats 13/65 hard · avg 18% 4.5% at its level · several runs
SciCode (level unspecified) (make… 34.7% █·····████●█ beats 5/6 very hard · avg 14% 3.9% at its level · several runs
chat & writing · #102 of 402 · 1.4% of its score
Beats [LongCat 2.0 (1.6T-A48B)](longcat-2-0.md), [Inkling Small (266B-A12B)](inkling-small.md), [GPT-5.4 nano](gpt-5-4-nano.md) here, though they rank higher overall. Trails [Mistral Large 3 (675B-A41B)](mistral-large-3.md), [Gemma 4 12B](gemma-4-12b.md), [Kimi K2 Instruct (0905)](kimi-k2-0905.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
IFEval (aggregation unspecified) … 93.0% ▁▂▄▂▂▄▆██●▂▁ beats 131/145 easy · avg 83% 1.4% maxed out · several runs
math · #93 of 431 · 23% of its score
Beats [Claude Sonnet 4.6](claude-sonnet-4-6.md), [Claude Opus 4.5](claude-opus-4-5.md), [MiniMax-M3](minimax-m3.md) here, though they rank higher overall. Trails [Gemini 2.5 Flash](gemini-2-5-flash.md), [Granite 4.2 30B](granite-4-2-30b.md), [Qwen3-Next 80B-A3B Instruct](qwen3-next-80b-a3b-instruct.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
HMMT February 2026 (maker) 81.2% █·▄··▄··▄█·● beats 7/7 hard · avg 32% 10% sharp · several runs
IMO-AnswerBench (maker) 74.2% ▄▄···▄·▄··█● beats 6/6 hard · avg 32% 7.1% at its level · several runs
AIME 2026 (maker) 90.7% ▁▁▁▂▂▂▆●▆▂·▁ beats 24/52 medium · avg 53% 6.1% sharp · several runs
web search · #2 of 15 · 4.8% of its score
test result vs the others in its run difficulty counts why
BrowseComp (maker) 40.9% ▅···█·▃···▃● beats 7/7 hard · avg 21% 3.2% at its level · several runs
Humanity's Last Exam with tools (… 21.8% ▁·▁▂●▁▄▄▇█▄▅ beats 8/62 very hard · avg 11% 1.7% at its level
other tests (5)
test result vs the others in its run difficulty counts why
Humanity's Last Exam (maker) 12.3% ▄▄·▄··▄·▄·●█ beats 5/7 very hard · avg 3.8% 11% several runs
GPQA (subset unspecified) (maker) 67.4% █···▃···●·▃▃ beats 5/7 hard · avg 37% 10% sharp · at its level · several runs
GPQA Diamond (maker) 73.7% ▁▁▁▁▂▄▅▅●██▄ beats 172/322 hard · avg 38% 7.9% sharp · at its level
LongBench v2 (maker) 48.7% ▂·▁·▂▃▁▄▄●█▄ beats 17/33 very hard · avg 32% 3.9% at its level
IFBench (aggregation unspecified)… 75.0% ▂▆▂▄▄▄▂▄▄▆●▇ beats 66/86 medium · avg 49% 2.7% at its level · several runs
source results share of the score
Spark-X2.6 model cards (submitted / mak… 17 54% ▇▇▇▇▇▇
XHToken/Spark-X2.5-4B model card (submi… 17 46% ▇▇▇▇▇
prior: pull toward the middle 0% ▇
weighted average = 64.7
Licences: Spark-X2.6 model cards ([Apache-2.0](../../licenses/spark-x26.txt)) · XHToken/Spark-X2.5-4B model card ([Apache-2.0](https://huggingface.co/XHToken/Spark-X2.5-4B/blob/0bcb35678590218655dff3765b9e61c83b35e9c4/README.md))
## How to run it
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run Spark-X2.5 4B (Q4_K_M, needs ~2.6 GB, from [huggingface.co/XHToken](https://huggingface.co/XHToken/Spark-X2.5-4B-GGUF/tree/9826e0be84e6e6e8b9668abc91421109a1df1e2d))
```sh
llama cli -hf XHToken/Spark-X2.5-4B-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf XHToken/Spark-X2.5-4B-GGUF:Q4_K_M
```
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run Spark-X2.5 4B (Q4_K_M, needs ~2.6 GB, from [huggingface.co/XHToken](https://huggingface.co/XHToken/Spark-X2.5-4B-GGUF/tree/9826e0be84e6e6e8b9668abc91421109a1df1e2d))
```sh
lms get https://huggingface.co/XHToken/Spark-X2.5-4B-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: 4bit (~2.3 GB, Mac only) `lms get https://huggingface.co/abenzerps/Spark-X2.5-4B-MLX-4bit -y` · 8bit (~4.4 GB, Mac only) `lms get https://huggingface.co/abenzerps/Spark-X2.5-4B-MLX-8bit -y`
### Ollama
1. Install Ollama. [Download it (macOS / Windows)](https://ollama.com/download)
```sh
# Linux
curl -fsSL https://ollama.com/install.sh | sh
```
2. Run Spark-X2.5 4B (BF16, needs ~8.2 GB, from [ollama.com](https://ollama.com/SparkLLM/Spark-X2.5-4B/tags))
```sh
ollama run SparkLLM/Spark-X2.5-4B
```
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run Spark-X2.5 4B (4bit, needs ~2.3 GB, Mac only, from [huggingface.co/abenzerps](https://huggingface.co/abenzerps/Spark-X2.5-4B-MLX-4bit/tree/f8fddb2c745bf14a8c28a47875d3d741721e84cf))
```sh
mlx_lm.chat --model abenzerps/Spark-X2.5-4B-MLX-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model abenzerps/Spark-X2.5-4B-MLX-4bit
```
Other versions: 8bit (~4.4 GB, Mac only) `mlx_lm.chat --model abenzerps/Spark-X2.5-4B-MLX-8bit`
Model repository and documentation: [huggingface.co/XHToken](https://huggingface.co/XHToken/Spark-X2.5-4B)
## Published speed measurements
Reference results for the hardware, engine and settings below. These are not verified speeds of the downloads above or predictions for your machine.
hardware precision speed engine and settings
AMD Radeon 890M (unified me… GSQ-RCO 3bit (p… 29.66 tok/s llama.cpp 4c9233c034fc, Vulkan · …
AMD Radeon 890M (unified me… GSQ-RCO 3.5bit … 23.32 tok/s llama.cpp 4c9233c034fc, Vulkan · …
NVIDIA H200 GSQ-RCO 3bit (p… 307.76 tok/s llama.cpp 4c9233c034fc, CUDA · ll…
NVIDIA H200 GSQ-RCO 3.5bit … 250.69 tok/s llama.cpp 4c9233c034fc, CUDA · ll…
NVIDIA GeForce RTX 3090 GSQ-RCO 3.5bit … 167.49 tok/s llama.cpp 4c9233c034fc, CUDA · ll…
NVIDIA GeForce RTX 3090 GSQ-RCO 3bit (p… 215.96 tok/s llama.cpp 4c9233c034fc, CUDA · ll…
NVIDIA GeForce RTX 2080 GSQ-RCO 3.5bit … 103.75 tok/s llama.cpp 4c9233c034fc, CUDA · ll…
NVIDIA GeForce RTX 2080 GSQ-RCO 3bit (p… 123.69 tok/s llama.cpp 4c9233c034fc, CUDA · ll…
AMD Radeon 890M (unified me… GSQ-RCO 3bit (p… 23.59 tok/s llama.cpp 4c9233c034fc, Vulkan · …
AMD Radeon 890M (unified me… GSQ-RCO 3.5bit … 16.02 tok/s llama.cpp 4c9233c034fc, Vulkan · …
Each speed links to its source, which has the full settings.
[interactive version](https://whatmodel.sh/model/spark-x2-5-4b) · [JSON](https://whatmodel.sh/api?format=json&all=1)
