Add SIQ-1-35B MoE — TrueScore 87.3 (#7 leaderboard)

[AlexWortega]

Summary

• SIQ-1-35B scores 87.3 TrueScore (rank #7), placing above DeepSeek V4 Flash and below Gemma 4 E4B on the leaderboard
• Canonical run: think-OFF, repeats=2, temp=0.3, Q4_K_M on single RTX A6000 46 GB
• Adds deployment recipe recipes/siq1-35b-moe.md with full serving command and results

Results

TrueScore	Quality	Cal	Rel	Eff	Resp	Pass@1	Pass@K	RelGap	Latency
87.3	79.2	88.6	93.0	100.0	96.8	91.2%	86.0%	5.3%	0.65s

Architecture note

SIQ-1-35B uses Qwen3_5MoeForCausalLM — a hybrid linear-attention MoE (3 linear-attn : 1 full-attn per block, 256 experts, 3B active params). vLLM and SGLang do not support this text-only arch — only llama.cpp serves it correctly via GGUF arch detection. Serving command:

llama-server -m SIQ-1-35B.Q4_K_M.gguf -ngl 99 -c 65536 \
  --host 0.0.0.0 --port 8080 --jinja --flash-attn on -np 1

Decode: 141 tok/s · Prefill: 4722 tok/s (linear-attn layers make long-context prefill very fast)

Test plan

• Ran canonical eval (repeats=2, thinking=off, temp=0.3, timeout=300)
• Run MD saved to results/runs/siq1-35b-20260626-134736.md
• Leaderboard table renumbered (rows 8–17)
• Serving method row added for SIQ-1-35B
• Recipe file added with full results table

🤖 Generated with Claude Code