[mle]/DaBench
Leaderboard

Spark X2.6 4B

XHToken/Spark-X2.6-4B-GGUF

LocalBF169.4 GB at 32K136 tok/s on RTX 5090605K context
Run recordCompareOpen weights
Overall score
15

17 of 153 test cases fully passed

Watch the test

Spark X2.6 4B Local Test | Beats Qwen3.5 4B? | Agentic Work, Coding & RAG Test | 16GB Local AI Setup

Watch on YouTube

What it built

What it wrote for showcase prompts, running live in your browser.

Scores by task

Scores give partial credit. Passed counts fully correct answers only. Select a task to read its test cases.

  1. 14Agents2 of 21 passedp90 latency 3m 37s
  2. 10Structured Extraction2 of 21 passedp90 latency 57.0s
  3. 4Coding0 of 21 passedp90 latency 1m 33s
  4. 10Grounded Answers2 of 21 passedp90 latency 1m 33s
  5. 37Instruction Following5 of 21 passedp90 latency 2m 20s
  6. 10Business Logic2 of 21 passedp90 latency 36.7s
  7. 31Security & permissions4 of 21 passedp90 latency 2m 4s
  8. 0Diagnosis & investigation0 of 6 passedp90 latency 2m 22s

Run it yourself

The exact build behind this score, and the command that serves it.

Deployment
Local
Engine
llama.cpp
Quantization
BF16
Server flags
-ngl -1 --temp 1.0 --top-p 0.95 --top-k 0 --reasoning on
Context window
605K
Memory
9.4 GB at 32K
Machine
RTX 5090, 32 GB
Throughput
136 tok/s
Cost of a suite run
$1.61est.
Cost per case
$0.011est.
Time per case
1m 3s
Measured
2026-10-21
Run this buildllama.cpp
llama-server -hf XHToken/Spark-X2.6-4B-GGUF:BF16 --port 9931 -c 619520 -ngl -1 --temp 1.0 --top-p 0.95 --top-k 0 --reasoning on

Pulls the GGUF from Hugging Face and serves an OpenAI-compatible endpoint on port 9931.

Memory it needs

7.7 GB of weights, 492 MB of runtime, and 39 MB per 1K tokens of context.

By context length
ContextMemorySmallest MacSmallest GPU
8K8.5 GB16 GB12 GB
32K9.4 GB16 GB12 GB
128K13.0 GB24 GB16 GB
1M47.1 GB96 GB80 GB

Estimated from the model's config. One slot, a 16-bit KV cache, no vision projector. A Mac fits when the build stays within the share of RAM macOS lets the GPU use, about two thirds up to 32 GB and three quarters from 36 GB. Macs and GPUs both keep 1 GB free for llama.cpp.

One email when a new model lands on the board or in the showcase. Unsubscribe anytime.