Ornith-1.5-9B

ADDED 2026-10-18. Open-weight dense reasoning model, the smallest member of the Ornith-1.5 family, built by continued pretraining and self-improvement RL on top of Ornith-1.0, which the card says was itself developed on top of Qwen3.5 and Gemma 4. The config model_type is qwen3_5 (Qwen3_5ForConditionalGeneration) and there is no num_experts key, so it is dense. FITS IN LOW VRAM: HF API safetensors total is 9,653,104,368 BF16, about 19 GB, which is the card's own size figure and small enough that the whole model sits in VRAM rather than spilling into RAM. A dense model that spilled would decode at RAM bandwidth and would not qualify; this one does not spill. PARAMS: params_b and active_params_b are both 9.653 from the safetensors count. flops_per_token_active = 2 x 9.653e9 = 1.9306e10 is the schema one-pass approximation. GEOMETRY from config.json fetched 2026-10-18: hidden 4096, 32 layers, 16 query heads, 4 kv heads, head_dim 256, so gqa_ratio 4. intermediate_size 12288, silu, vocab 248320, tie_word_embeddings false, partial_rotary_factor 0.25. layer_types is 24 linear_attention + 8 full_attention, one every fourth layer (indices 3, 7, 11, 15, 19, 23, 27, 31). KV (bf16): only the 8 full-attention layers grow with context, 8 x 4 kv heads x 256 x 2 (K+V) x 2 bytes = 32,768 B/token. The 24 linear-attention layers hold a fixed recurrent state and are not in that figure. If all 32 layers cached the full context it would be 131,072 B/token, 4x the recorded number. No rope_scaling key and mrope interleaved, so context_scaling_method is mrope. The card's own recipe serves BF16 on one 80 GB GPU (vLLM --tensor-parallel-size 1, gpu-memory-utilization 0.90, 262144 context). BENCHMARKS ARE VENDOR CLAIMS, averaged over five runs on the card. BEATS models/nemotron-labs-dory1-4b, the catalog's best sub-10B model with a published SWE-bench Verified number, on SWE-bench Verified: 70.6 vs 61.8. Its GPQA Diamond 86.4 also beats that record's best four-loop vendor claim of 72.16, and its Terminal-Bench 2.1 (Terminus-2) 46.2 is far above the 8.6 vendor figure on models/minicpm5-2b. It does not beat models/xing4-0-29b-a4b on SWE-bench Verified (70.6 vs 75.00) or on Terminal-Bench 2.1 (46.2 vs 57.50); Xing is a 4B-active MoE in a larger class. The card's comparison rows for Qwen3.5-9B (SWE-bench Verified 53.2, GPQA Diamond 82.8, Terminal-Bench 2.1 21.3) and Qwen3.6-35B-A3B are not catalog records. The HF eval-results widget repeats SWE-bench Verified 70.6 and GPQA Diamond 86.4.