A user benchmarks local agent models on dual RTX 5070 Tis, finding that Ornith 1.5 35B-A3B (a ~3B-active-parameter MoE) matches Qwen3.8 27B on their tests while running about 3x faster (~180 tok/s) and scaling to 512K/1M contexts with YaRN.
My setup is two RTX 5070 Ti 16GB cards (the second one is on an OCuLink dock) with 64GB of RAM, Ollama on Windows, and the agent runs on pi in WSL. Until last night the daily model was Qwen3.8 27B UD-Q4_K_XL at 128K with MTP, which does about 55 to 70 tok/s across both cards. I have a weekly job that looks for new open models and runs anything that fits through two tests I built for my agent. One is a 9 step long session (tool calls, reading files, a decision, and recall after the context compacts three times). The other is 10 small coding tasks. The 27B gets 9/9 and 10/10. This week it picked up Ornith 1.5 35B-A3B (ornith-1.5:35b in the Ollama library, Q4_K_M). It passed 9/9 and 10/10. Laguna XS 2.1 also passed both. North Mini Code 1.0 only got 4/9. Ornith at 128K context is 24.4GB and sits fully on the two cards. Generation is about 180 tok/s (176 and 183 on two runs, short prompt, thinking off). That's around 3x what the 27B gave me. The speed makes sense once you look at the model info. Only about 3B params are active per token (256 experts, 8 used), and only 10 of the 41 layers are full attention, with 2 KV heads. The rest are linear attention, so the KV cache barely grows. Going from 128K to 256K only added about 2GB. 256K does fit, but about 1.2GB ends up in system RAM because my first card also runs the monitors, so it drops to about 139 tok/s. I left 128K as the default and made 256K something I switch to when I need it. Caveats: both of my tests max out, so this only shows it isn't worse than the 27B on my workload. It doesn't prove it's smarter. Artificial Analysis hasn't scored it yet. The vendor numbers are 79 on SWE-bench Verified and 68.5 on Terminal-Bench 2.1, which I haven't checked myself. Next I'm trying 512K and 1M on llama-server. The model card says YaRN at factor 4 on top of the native 262144 gets you about 1M, and factor 2 about 512K. I'll post numbers if it holds up. Anyone else running it for agent work? Curious how it does for you on long sessions compared to the 27B. Edit: the long context runs held up. On llama-server with YaRN set the way the model card says, 512K (factor 2, q8 KV cache) fits fully on the two cards and found a note I planted about 335K tokens into a 419K token prompt. It read that at about 1060 tok/s on average and generated about 35 tok/s at that depth. 1M (factor 4, q4 KV cache) only loaded once I let llama-server's fit option push some experts to system RAM, and it found the note at about 720K in an 849K prompt. That one took about 24 minutes to read (570 tok/s average) and generated about 16 tok/s. On short prompts it's about 135 tok/s at 512K and about 68 at 1M.
User runs local benchmarks comparing Qwen3.6 27b, Gemma4 26B, and Ornith1.0 35B on an RTX 3090 using inspect-ai. Results show Qwen leading in knowledge and coding, while Ornith is competitive in grounding and recall.
A user reports achieving 60 tokens per second with the Ornith-1.5-35B-A3B MOE model on an NVIDIA RTX 4070 Ti, demonstrating efficient local inference optimization.
The post details running the Ornith-1.5-9B AI model on an RTX 3060 with 200k context, achieving high inference speeds using advanced quantization techniques.
User tests Ornith-1.0-35B on an RTX 3090, finding fast inference speeds (1560 tok/s prompt, 78 tok/s generation) but consistently worse coding performance on Three.js tasks compared to Qwen 3.6, even after multiple attempts.
A user reports that Ornith-1.0-35B matches Qwen3.6-35B in performance but excels at planning and long task execution, while the developer announces the open-source Ornith-1.0 family of LLMs specialized for agentic coding.