3-Way MoE Model Comparison

K2 Horizon 36B A4B  vs  Qwen 3.6 35B A3B  vs  Ornith 1.5 35B A3B — 8 benchmarks + throughput

K2 Horizon Wins
4
Avg: 82.4
Ornith 1.5 Wins
3
Avg: 82.4
Qwen 3.6 Wins
1
Avg: 73.4
Throughput (1x / 8x)
88 / 672
Qwen fastest single, Ornith wins batch
K2 Horizon 36B A4B
36B total / 4B active • MoVA • 512K context • oQ4e
Qwen 3.6 35B A3B
35B total / 3B active • Hybrid DeltaNet+Attn • 1M context • oQ4e-mtp
Ornith 1.5 35B A3B
35B total / 3B active • Hybrid 30L+10A • 262K context • oQ4e-mtp

Overall Performance Profile

Benchmark Scores

Individual Benchmark Detail

Throughput Comparison (oMLX)

Single Stream (4K prompt)

Batch Scaling (8x)

Verdict by Dimension

Quality (Avg Score)

K2 Horizon / Ornith 1.5 (tie)

Both average 82.4. K2 wins 4 benchmarks, Ornith wins 3. Different strengths.

Throughput (Single)

Qwen 3.6

88.4 TG TPS at 1x. ~1.8x faster than K2. MTP + 3B active advantage.

Throughput (Batch)

Ornith 1.5

671.6 TG TPS at 8x batch. Beats Qwen (648) and K2 (354). Best scaling.

Tool Calling

Ornith 1.5

Perfect 100/100. Sweeps all 5 categories. K2 close at 93.

Safety

K2 Horizon

0 harmful violations. Perfect on Hard System vs Tool Data. PA=80.

Agentic (Hermes)

Ornith 1.5

85/100 on Hermes agent suite. Perfect Memory & Recall. Strong delegation.

Multimodal

K2 Horizon / Qwen 3.6

Qwen: text+image+video. K2: text-only. Ornith: text+image.