Ornith1.5 35B (3B active)

Source: https://github.com/lazarevtill/strix-halo-llm/blob/main/docs/BENCHMARKS.md#b11414-speculation-at-the-serving-temperature-and-new-models-2026-10-05 Author-reported native llama-bench campaign dated 2026-10-05, reconstructed and source-checked 2026-10-13. AMD Ryzen AI MAX+ 395 / Radeon 8060S (gfx1151, Strix Halo), one integrated GPU, 128 GB unified/shared system memory; Windows 11. Hardware/platform attribution: https://github.com/lazarevtill/strix-halo-llm Model: Ornith-1.5-35B-A3B, Q6_K GGUF, MoE with approximately 3B active parameters. 35B is the nominal model class, not an exact counted total; the source download registry calls it 36B/A3B. Model identity: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B Source-attributed measured GGUF repository/file: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF/blob/main/Ornith-1.5-35B-Q6_K.gguf The source's ornith15 download mapping explicitly identifies this first-party repository and filename: https://github.com/lazarevtill/strix-halo-llm/blob/main/scripts/windows/fetch-models.ps1 The tested HF revision and file checksum were not reported. The linked main file is the current artifact location, not a claim of an immutable tested revision. Source registry lists 29,208,731,392 bytes; file size is not peak resident memory. Runtime: llama.cpp b11414, Vulkan. The table is an engine A/B against b11330, with solo occupancy, same session, microbatch 1024, and 2 repetitions. Batch 2048 is campaign-linked through the surrounding benchmark documentation and harness, rather than a recovered complete command for this exact run. All five b11414 measurements, in tokens/second: - pp512, empty-cache/depth-0 row: 999.77 +/- 14.55 - pp4096, empty-cache/depth-0 row: 1001.92 +/- 5.23 - tg128, depth-0 row: 57.34 +/- 0.16 - pp4096 at 32,768 existing KV tokens: 563.23 +/- 4.53 - tg128 at 32,768 existing KV tokens: 50.14 +/- 0.09 Headline PP 563.23 and TG 50.14 select the two separate depth-32768 tests. PP adds 4096 tokens to an existing 32768-token cache; TG measures generation of 128 tokens at that starting depth. They are separate synthetic tests, not one timed 4096-in/128-out request, total request speed, or an end-to-end wall-clock rate. Existing KV depth is not the configured context limit. The source preserves its +/- values; their precise statistical definition is not restated here. Timer provenance: the published table belongs to the repository's llama-bench speed measurements. The public bench-big.ps1 parses the test and t/s columns from native llama-bench stdout; its outer process Stopwatch supplies run-duration/status messages, not these PP/TG rates: https://github.com/lazarevtill/strix-halo-llm/blob/main/scripts/windows/bench-big.ps1 That currently visible harness uses -ngl 999, -fa 1, -ctk q8_0, -ctv q8_0, -lm none and -o md. These corroborate how its speed harness works but are not proof of every flag in the historical b11414 invocation. No full invocation or native stdout file for this exact table was recovered. b11330 is a separate same-session control, not combined into this result. The neighboring DFlash/MTP, temperature-0.6, seed-42, n_predict-256, server-prefill and multi-slot serving experiments are separate measurements and do not define this native campaign. Configured context capacity, exact engine commit, GPU driver, power/thermal settings, exact historical cache precision/offload flags, and peak resident RAM/VRAM for these five rows remain unreported/unverified. Do not import the source's serving context 2 x 262144 or serving-memory figures.