Ornith 1.5 35B Is Smarter, But Is It Worth the Slowdown?

Описание: Ornith 1.5 posts stronger published agentic-coding scores than Ornith 1.0, but the new release is not the local speed winner. Is the capability tradeoff worth it?

I tested Ornith 1.5 35B Q4_K_M against Ornith 1.0 35B Q4_K_M and Qwen 3.6 35B UD-IQ4_NL in llama.cpp on the same Ryzen 7 and RTX 4070 Ti Super system.

Published agentic-coding context

NL2Repo: Ornith 1.5 scores 46.2 versus Ornith 1.0 at 34.6, a 11.6-point lead.
SWE-bench Verified: 79.0 versus 75.6.
SWE-bench Pro: 59.6 versus 50.4.
SWE-bench Multilingual: 71.4 versus 69.3.
ClawEval: 72.5 versus 69.8.

Those are vendor-published model-card results, not locally reproduced task-quality scores. The thumbnail columns depict the published NL2Repo comparison.

Local llama.cpp results

Ornith 1.0 was the fastest decoder at 77.28 tokens per second.
Ornith 1.5 generated at 74.64 tokens per second, 3.4 percent behind Ornith 1.0.
Qwen 3.6 processed the mixed-context workload at 1,309.22 tokens per second, 45.4 percent ahead of Ornith 1.5's 900.38 tokens per second.
At 77K context, Qwen remained the prompt-ingestion winner while both Ornith models decoded faster.

Test setup

98,304-token context
Flash Attention enabled
q8_0 KV cache
11 uncached prompts from 30 to 77,376 tokens
256 generated tokens per prompt
No multi-token prediction tested

Important caveat

The Ornith models used Q4_K_M while Qwen used UD-IQ4_NL. This is a same-machine deployment comparison, not a quantization-controlled architecture study. The near-identical VRAM snapshots also do not prove equal memory cost or complete GPU residency.

Sources:
https://huggingface.co/ornith-ai/Orni...
https://huggingface.co/ornith-ai/Orni...

#LocalLLM #llamacpp #Ornith #Qwen #llm