VeriLoop E2

ADDED 2026-10-18. Fine-tune of Qwen/Qwen3.8-27B by Tsinghua SIGS Robot Lab, weights public 2026-09-22 (HF API createdAt 2026-09-22T14:23:55Z) and the card last touched 2026-10-10. RUNNABLE AS DENSE: this is the same 27.781B checkpoint size as models/qwen3-8-27b, which the catalog already records as the runnable dense class, so the whole model fits where that one fits and nothing is offloaded to RAM. config.json fetched 2026-10-18 has no num_experts key, so active_params_b equals params_b. The only config addition over the base is a top-level veriloop_conditional_memory block; the text geometry is unchanged. BEATS models/step-5-preview, the catalog's highest recorded GPQA Diamond number, on GPQA Diamond: 93.9 vs 93.5. The same card also beats models/beam-1-0-501b-a23b on AIME 2026 (98.3 vs 97.7) and models/ling-3-1-flash on SWE-bench Pro (76.2 vs 68.74). All three of this card's numbers are vendor claims from .eval_results YAML rows that point at the VeriLoop-E2-Evaluation-Evidence dataset and carry no verified flag and no harness description, so the comparison is across harnesses and is not independently re-run here. Step 5 Preview and Beam are large-total MoE with a far smaller active footprint; this one is a 27.8B dense hybrid, so the win is on the published number, not on offload cost. Terminal-Bench 2.1 at 88.8 beats models/qwen3-8-27b (73.0) and models/mimo-v2-6-flash-rl (87.6) but does NOT beat models/mimo-v2-6-pro-rl (89.9), so Terminal-Bench is not the qualifying comparison. SWE-bench Verified at 81.4 does not beat models/beam-1-0-501b-a23b (81.6), and LiveCodeBench v6 at 84.7 does not beat models/kimi-k2-6 (89.6). PARAMS: safetensors BF16 count 27,781,427,952, identical to models/qwen3-8-27b, so params_b and active_params_b are both 27.781. The card rounds this to 27B. GEOMETRY from config.json text_config, fetched 2026-10-18: hidden 5120, 64 layers, layer_types is 48 linear_attention + 16 full_attention at full_attention_interval 4 (indices 3,7,...,63), num_attention_heads 24, num_key_value_heads 4, head_dim 256 so gqa_ratio 6, intermediate 17408, silu, vocab 248320, tie_word_embeddings false, max_position_embeddings 262144. Linear block: linear_num_key_heads 16, linear_num_value_heads 48, key and value head dim 128, conv kernel 4. RoPE is native mrope: rope_type default, mrope_interleaved true, mrope_section [11, 11, 10], partial_rotary_factor 0.25, rope_theta 10000000, no yarn/ntk factor. context_scaling_method is mrope. KV (bf16, full-attention layers only): 16 x 4 kv heads x 256 x 2 (K and V) x 2 bytes = 65,536 B/token (64 KiB), the same derivation as models/qwen3-8-27b. The 48 linear-attention layers hold a fixed recurrent state, not a growing cache. flops_per_token_active = 2 x 27.781427952e9 = 5.556e10, the schema one-pass approximation; the linear layers are not 2 FLOPs per parameter, so treat it as the schema convention, not a profile. Vision tower is separate: vision_config depth 27, hidden 1152.