v0.11.0

Decima, three small multilingual decision models; Winnow 12B reads images; and a thread count for CPU servers.

New models: Decima (#54)

Model	By	What it is	Measured
decima:base (decima)	A. M. Madani	mmBERT-base (321M) with a late-interaction scorer: the state and each option are encoded on their own and every option reads the state, so option order never changes the answer	0.495 on typed-decisions, ECE 0.073; 15 ms for five questions on an RTX 4090
decima:agent	A. M. Madani	decima-base fine-tuned for the decisions inside a coding agent's loop: secret and command gates, tool, command and model choice; states up to 2,048 tokens	0.486 on typed-decisions; 0.93 on the author's 130 agent decisions
decima:small	A. M. Madani	multilingual-e5-small (122M) with the same scorer	0.432 on typed-decisions; 146 ms for five questions on an i9-13900K, the fastest model Ollaya runs on a CPU

ollaya run decima --preset triage "My order never arrived and support ignores me. Refund me today or I'm switching to your competitor."

• Parity. Ollaya's Rust runtime matches the author's own model.py and systemone.py in fp32 on 581 questions per model, on the CPU, an RTX 4090 and an RTX 5090: identical token rows, the same decision on every question, scores within 2.0e-5 and probabilities within 5.5e-6.
• Weights. The author's fp32 PyTorch checkpoints (pytorch/ of each repository), pulled unmodified from Hugging Face, pinned to a commit and checked by sha256. Ollaya hosts only the 3 to 5 MB graphs.
• Limits. States of 512 tokens per question (2,048 for agent); /v1/systemone answers a longer one with STATE_TRUNCATED, as the author's server does. 2 to 255 choices, 2 to 10 score levels.

Winnow 12B reads images (#59)

winnow:12b-vision is winnow:12b plus the author's matching vision projector, read through libmtmd as winnow:e4b-vision is since 0.10.0. PNGs go in images on /api/decide or --image on the command line. Parity against stock llama-server b11146 on the same device: 65 of 65 image questions on an x86-64 CPU and an RTX 5090, option logits within 7.7e-6. It peaks at about 16 GB on the GPU. Thanks @Txoka.

OLLAYA_THREADS (#60)

The server now passes a thread count to every model it loads: ONNX Runtime's intra-op threads, or llama.cpp's threads for GGUF models. It matters most on CPU servers without SMT, where llama.cpp's default (half the CPUs on x86-64) left half the machine idle: on a 6-vCPU VPS, OLLAYA_THREADS=6 took winnow:e4b from about 96 s to a median of 67 s on a long request. Unset keeps each engine's default. Thanks @hkulekci.

Upgrade

ollaya update

or curl -fsSL https://ollaya.dev/install.sh | sh. Decima and winnow:12b-vision need 0.11.0 or newer.

What's Changed

• Add opt-in Winnow 12B image decisions by @Txoka in #59
• OLLAYA_THREADS: CPU threads for each loaded model by @hkulekci in #60

Full Changelog: v0.10.0...v0.11.0