External Publication
Visit Post

Best MoA with models under 40B

Hugging Face Forums [Unofficial] July 3, 2026
Source

This combination is doable with total of around 90GB of RAM and it is superior to DeepSeek-v4-pro Please see my other post. Any one can predict the combination: Drafters:

  1. Gemma-4-31B-QAT-Q4 with MTP (llama.cpp)
  2. Ornith-1.0-35B - Q4_k_M (llama.cpp) each with a context of 102400 and 8bit kv cache Aggregator: Qwen3.6-27B-4bit context 204800

Call it MoA-Max:

Task MoA-MAX
Bug Investigation 89.2
Tool Orchestration 89.3
Code Architecture 87.8
Clinical Decision 80.4
Multi-step Coding (DLQ) 91.8
Overall 87.7

Compared to single and other MoA:

Rank Model / Configuration Bug Tool Arch Clinical DLQ Avg /100
1 MoA-MAX † 89.2 89.3 87.8 80.4 91.8 87.7
2 MoA: Gemma-12B + Ornith-35B → agg Qwen-3.6-27B † 80.3 86.6 88.1 90.3 88.9 86.8
3 Qwen-3.6-27B (4-bit, vLLM) 90.8 86.2 86.2 80.7 88.9 86.6
4 MoA: Gemma-12B + Qwen-27B → agg DeepSeek-v4-Pro 88.1 88.1 88.1 81.1 84.2 85.9
5 Qwen3.6-35B-A3B (Q4_K_M) 88.1 83.1 88.1 77.6 88.9 85.2
6 MoA: Gemma-12B + Nemotron-2 → agg Qwen-3.6-27B 80.5 87.5 88.1 80.5 88.9 85.1
7 MoA: Gemma-12B + Qwen-35B-A3B → agg Qwen-3.6-27B 81.1 88.1 88.1 77.6 88.9 84.8
8 MoA: Gemma-12B + Qwen-27B → agg Qwen3.6-35B-A3B 88.1 88.1 86.6 81.1 77.6 84.3
9 DeepSeek-v4-Pro (cloud) 93.2 88.1 88.1 55.2 94.0 83.7
10 DeepSeek-v4-Flash (cloud) 79.5 88.1 88.1 73.6 84.9 82.8
11 MoA: Gemma-12B + Qwen-27B → agg Ornith-35B 80.5 82.4 84.7 81.1 84.9 82.7
12 Ornith1.0-35B (Q4_K_M) 73.0 88.1 86.2 81.1 84.9 82.7
13 MoA: Gemma-12B + Qwen-27B → agg Nemotron-2 59.1 61.5 78.2 61.9 72.9 66.7
14 Gemma-4-12B (4-bit, vLLM) 69.5 51.5 56.1 43.9 49.5 54.1
15 Nemotron-2 Cascade (Q4_K_M) 5.8 5.8 5.8 5.8 5.8 5.8

See my other post for details

Discussion in the ATmosphere

Loading comments...