ai-tier / edge

K2 Horizon 3.7B on Pi 5

Snapshot · ai-tier

Raspberry Pi 5 · 16 GB RAMAll performance measurements are for this device.

Capability shows how well the model answers the benchmark questions. Speed, response time and memory are measured on Raspberry Pi 5 · 16 GB.

Capability summarises instruction following, tool calling and finding information in long prompts. Speed and memory refer to the Raspberry Pi 5. — means unavailable.
ModelCapability / 100Output tok/sPrompt tok/sFirst token · sTask time · sPeak memory · MiBParameters · B
K2 Horizon 3.7BQ4_K_M · 40,960 context · reasoning on653.921.119.2144.810,3565.1

Benchmark breakdown

Instruction following · IFEvalTool calling · BFCLFinding information · RULER
735765

Test configuration

Model publisher
Institute of Foundation Models / MBZUAI
Model parameters
5.058 billion total
Published input support
text (metadata; not device validation)
Quantization
Q4_K_M
Context capacity
40,960 tokens
Reasoning mode
on
CPU threads / parallel requests
4 / 1
Temperature / seed
0 / 42
Measurement date
2026-10-21T21:03:52.923100+00:00
Result ID
split-5a57437c64a9770de8bd
Runtime revision
620d2225cc436ab0b001e3200e590e3b286ef9ab
Artifact repository / revision
IFM/K2-Horizon-3.7B-GGUF / 78a0635e8f19b880e468a2fc82923ee6c8a469e4
Artifact SHA256
07773aed93890d4f5d08b421430000d5c8d73cefc74fbdc32e67d820f23207b2
Dataset revision
a0fc00b6a07455d1efb70cc19d5ca94c5923c5261a0b41095083a41a1592a1f6
Configuration identity
918bf77ac3c4bfa79d2bff8f177378f23fcf02d2df257aed3b2cfb16f7036d90

Measurement limits

Speed tests read 256 input tokens and generate 128 output tokens, repeated ten times. The output speed test starts without a long conversation history. Answer time uses six fixed prompts with up to 128 output tokens, repeated three times. First-text time is measured from the application making the request. Peak memory includes the loaded model; it excludes other applications on the device. Longer prompts may need more memory. Capability uses a fixed selection of public questions, rather than the full official benchmarks.

Measured 2026-10-21T21:03:52.923100+00:00.