Ollaya is an Apache-2.0 local runtime and model hub, built in the image of Ollama, that serves twenty-nine open decision-model families behind a wire-identical replica of TypeSafe’s Jev API.
The decision-model wave just got its Ollama, and its verification discipline is the strongest the open side of this category has shown: every shipped model measured on the released runtime and checked answer by answer against its author’s own code, with the raw data published.
What it is #
One binary (a desktop app, a CLI, and a Docker image) for macOS, Windows, and Linux that exposes /v1/systemone, /v1/decisions, and /v1/models with TypeSafe’s request and response formats, so the official TypeSafe Python SDK 0.7.1 runs unchanged against localhost.
The library spans twenty-nine families as of 2026-10-26: encoder-based ones (laya, largha, nli, gliclass, von, decima, and Liquid AI’s d1-omni) that answer in 10 to 20 milliseconds on a CPU, decoder-based ones (winnow, clef, kev, decider, nimble, jeb, jeeves, cygnet, snap, arbiter, Perplexity’s pplx-decider, Liquid AI’s d1, Knowledgator’s gliclass-v4 decoder-KV line, and the vLLM Semantic Router’s decision and d3 lines) up to 27B, and Ollaya’s own first-party lines (coba, Laya’s network on Ettin encoders, and sura, Qwen3.5 decoders, both distilled from pplx-decider:flash).
Weights are pulled from each author’s Hugging Face repository pinned to a commit and checked against sha256, never re-hosted.
It runs on ONNX Runtime over the CPU or an NVIDIA GPU (CUDA 13 or 12, Vulkan), with MLX on Apple silicon for laya and nli, and it is an independent project by Mert Cobanov, not affiliated with Ollama or TypeSafe.
Status #
Thirty-three days old, beta, and already the wave’s default front door. The repository was created 2026-09-23 and shows 1,313 stars and 79 forks as of 2026-10-15, pushed 2026-10-25. The release train moved fast after the launch measurements: v0.9.0 (2026-10-02) packaged Cloudflare’s Clef and a CPU fallback for GGUF models, v0.10.0 (2026-10-05) added Hugging Face mirror pulls and a DGX Spark CUDA pack, v0.11.0 (2026-10-06) added Decima’s three small multilingual models and image reading for Winnow 12B, v0.12.0 (2026-10-06) shipped Vulkan-on-Windows GGUF support, v0.12.1 (2026-10-09) added a new Credence family (a MiCA refinement of Winnow-E4B), and v0.13.0 (2026-10-13) added Liquid AI’s d1 family (a 3B LFM2.5-VL decision model with a vision variant and the library’s first revenue-threshold license, the LFM Open License v1.0), moved Decima’s agent tag to 2.2, added digest-pinned pulls, and fixed the CUDA idle-pause crash that hit ONNX decoder models, at which point the site’s family count read twenty-one.
The launch Show HN thread (2026-09-25) reached 618 points, the third-largest of the whole Jev wave.
The measurement story moved on 2026-10-14: Ollaya Bench (bench.ollaya.dev, Edition 2026.10, built on Ollaya 0.13.0) supersedes the 0.8.0-era results page, scoring 36 models over 14 tasks (13 public datasets plus a 2,000-question typed-decisions set, 5,880 questions per model, now issued over HTTP with memory recorded beside latency) on two disclosed devices (an RTX 5090 and a Threadripper 3960X CPU board), with chance-corrected scores, ECE, median latency, VRAM, and license per model; jeb:27b leads at 67.8 (253 ms, 16.2 GB), jeb:4b is the sub-100-ms pick at 63.9, and the superseded page’s 0.8.0-era counters (31 models, 73,720 benchmark answers, 41,352 parity checks) remain its historical record.
The release train added v0.14.0 on 2026-10-19: fp16 decoder tags, a model lifecycle, and a refreshed human-labeled comparison now against Ollama 0.40.6 (nine models) whose new accuracy leader is pplx-decider:flash at 0.824 against Ollama’s best 0.816 (the 27B pplx-decider), with strands-decider and nimble:9b-v3 measured side by side, and the same-Nimble calibration gap now reading 5.1x (ECE 0.022 on Ollaya against 0.115 on Ollama) while Ollama stays five to eight times faster on the same weights.
The train continued with v0.15.0 (2026-10-20), which added the Largha family, surfaced laya:typed-decisions and jeb in the site’s typed-decisions scoreboard, and opened a Libraries section on the TypeSafe compatibility page, and v0.16.0 (2026-10-21), which added AMD ROCm/HIP support on Linux with a ROCm pack shipping in the release assets and an install.sh --uninstall, at which point the site’s family count read twenty-four.
The site’s count moved to twenty-six as the first-party lines shipped: coba (Laya’s network on Ettin encoders, distilled from pplx-decider:flash, its training report closing the records’ fourteen TODOs) and sura:0.8b and sura:2b (Qwen3.5 decoders distilled from pplx-decider:flash in its joint format, on the 2026-10-24 push), and v0.17.0 (2026-10-23) aligned run output when an answer fills the column, built the linux-amd64 release on Ubuntu 24.04, and made parity_llama fail when the goldens hold no questions.
Its head-to-head numbers are self-run but use Bespoke Labs’ public benchmark with Ollama on the same GPU: winnow:12b’s 0.773 remains the strongest non-pplx row.
v0.18.0 (2026-10-25) then added the d1-omni, gliclass-v4 (Knowledgator’s decoder-KV generation), and d3 (the vLLM Semantic Router’s Decision 3.0) families, moved jeeves:9b to Jeeves v2 and the decider tags to Mapika’s 2026-10-20 update, brought DeBERTa, Qwen3, Decision 1.0, and kev onto the Apple GPU under MLX, shipped parity smoke tiers (K1 through K3) and a redesigned site with a desktop-app release build, and closed with a first draft of the coba technical report for arXiv, at which point the site’s count read twenty-nine.
Strengths #
- Parity checking against the authors’ code is the discipline this category needed: answers compared one by one against what each model’s author ships, raw data included, rather than another self-graded accuracy table.
- Weights pinned by commit and sha256 from the authors’ own repositories, so the runtime cannot silently drift from what an author measured.
- Calibration is first-class: each model ships its author’s fitted temperature, and a Modelfile refits it on your labelled data.
- The encoder families (nli, gliclass, laya) cover the millisecond, CPU-friendly slice that Ollama’s decoder-only decision support ignores, and the TypeSafe-compatible surface means code written against the closed API moves without edits.
Cautions #
- The accuracy and latency comparisons are self-run on two disclosed consumer machines, and the hosted-Jev comparison imports third-party benchmark numbers measured elsewhere, which the site itself says to read only as orders of magnitude.
- Beta software with one named developer carrying it, so bus factor is one.
- The library mixes strong and weak models by design (its own table scores laya:en at 0.361 on typed decisions), so the runtime does not save you from model choice.
- API compatibility is to the published formats, not to the closed model: point the TypeSafe SDK at Ollaya and you get open-model quality, not Jev’s.
Pricing #
Free and open: Apache-2.0 runtime, authors’ weights under their own licenses, no metering and no hosted tier. The cost is your hardware and the model downloads.
Compared to #
- Jev: the closed, hosted original Ollaya’s API copies; choose Jev for frontier accuracy behind an SLA, Ollaya for local, private, unmetered decisions across twenty-nine families.
- Kev: a single-family server for the same contract; Ollaya is the multi-family hub that ships kev:9b among its models.
- Ollama 0.35: the incumbent runtime, which serves two decision-model families (Nimble and Tev1) as raw-softmax GGUF conversions; Ollaya covers more families with author calibration but loses the same-weights latency race.
Bottom line #
Recommended for engineers who want the open decision-model wave behind one local API with its calibration intact, and for anyone auditing a checkpoint against its author’s code. Not for anyone who needs a vendor SLA, hosted-Jev-class accuracy today, or a project with more than one maintainer. The disagreeable claim I will defend: runtimes, not models, decide which open ecosystems survive, and Ollama proved it for LLMs, so every model author in this wave should treat Ollaya’s defaults as their distribution channel whether they like it or not.
Changes #
- 2026-10-06 - Created from the entrant scan after the 2026-09-25 Show HN thread cleared the bar (618 points, 1.2k stars in ten days, and a parity-checked results page).
- 2026-10-06 - Corrected the age claim (created 2026-09-23, thirteen days old, not three weeks) and refreshed stars to 1,209.
- 2026-10-07 - Added the ollaya-dev/ollaya star history chart to the Status section.
- 2026-10-07 - Reworded two pre-existing “shapes” compounds to “formats” (the TypeSafe request and response formats, the published formats) to keep the banned-terms rule.
- 2026-10-08 - The library grew to nineteen families (decima, snap, and arbiter joining) and the results page moved to 31 models measured, 2,184 timed requests, a newest measurement of 2026-10-06, a third disclosed machine (M4 Pro Mac mini) beside the two desktops, and Ollaya 0.8.0; refreshed stars to 1,239.
- 2026-10-12 - Recorded the release train v0.9.0 through v0.12.1 (2026-10-02 to 10-09): Clef and a GGUF CPU fallback (0.9.0), Hugging Face mirror pulls and the DGX Spark CUDA pack (0.10.0), Decima’s multilingual trio and Winnow 12B image reading (0.11.0), Vulkan-on-Windows GGUF (0.12.0), and the new Credence family (0.12.1), while the site’s family count and results page still carry the 0.8.0-era state; refreshed stars to 1,298.
- 2026-10-13 - Recorded v0.13.0 (2026-10-13): Liquid AI’s d1 family joined (a 3B LFM2.5-VL decision model with a vision variant, the library’s first revenue-threshold license in the LFM Open License v1.0, and Ollaya’s own parity-and-accuracy measurements), Decima’s agent tag moved to 2.2, pulls can pin a manifest digest, and the CUDA idle-pause crash affecting ONNX decoders (kev, decider, decision, nimble, jeeves, clef) is fixed; the site’s family count moved to twenty-one; refreshed stars to 1,301 and the pushed date to 2026-10-13.
- 2026-10-17 - Volatile re-verification after the 2026-10-14 and 10-17 pushes: recorded Ollaya Bench (bench.ollaya.dev, Edition 2026.10, 2026-10-14, on 0.13.0), the new measurement site superseding the 0.8.0-era results page (36 models, 14 tasks, 5,880 questions per model over HTTP, two disclosed devices, jeb:27b the leader at 67.8), and reworked the thesis and status around it; pushed date refreshed to 2026-10-17.
- 2026-10-20 - Volatile re-verification after the 2026-10-19/20 pushes: recorded v0.14.0 (2026-10-19, fp16 decoder tags and a model lifecycle), the library grown to twenty-three families with Perplexity’s pplx-decider joining, and the human-labeled comparison re-based on Ollama 0.40.6 with pplx-decider:flash the new accuracy leader at 0.824 (Ollama’s best 0.816) and the Nimble calibration gap now 5.1x (0.022 against 0.115); pushed date refreshed to 2026-10-20.
- 2026-10-22 - Volatile re-verification after the 2026-10-22 push: recorded v0.15.0 (2026-10-20, the Largha family, laya:typed-decisions and jeb in the typed-decisions scoreboard, a Libraries section on the TypeSafe page) and v0.16.0 (2026-10-21, AMD ROCm/HIP support on Linux with a ROCm pack in the release assets and install.sh –uninstall); the library now twenty-four families per the site; pushed date refreshed to 2026-10-22.
- 2026-10-24 - Volatile re-verification after the 2026-10-23/24 pushes: recorded v0.17.0 (2026-10-23, run-output alignment, the linux-amd64 release built on Ubuntu 24.04, parity_llama failing on empty goldens) and the library grown to twenty-six families with Ollaya’s first-party lines joining (coba on Ettin encoders and the sura:0.8b/2b Qwen3.5 decoders, both distilled from pplx-decider:flash); pushed date refreshed to 2026-10-24.
- 2026-10-26 - Volatile re-verification after the 2026-10-25 push and v0.18.0: recorded the d1-omni, gliclass-v4, and d3 families joining (the site’s count to twenty-nine), jeeves:9b moved to Jeeves v2 and the decider tags to Mapika’s 2026-10-20 update, the MLX phases putting DeBERTa, Qwen3, Decision 1.0, and kev on the Apple GPU, parity smoke tiers K1 through K3, the redesigned site with a desktop-app release build, and the coba technical report’s first arXiv draft; pushed date refreshed to 2026-10-25.
See also #
- Jev - the closed model whose API Ollaya replicates wire for wire
- Kev - the single-family server whose checkpoints Ollaya also packages
- Laya - the encoder family that gives Ollaya its CPU-speed slice
- Nimble - whose public human-labeled benchmark is the yardstick in Ollaya’s Ollama comparison
- Hybrid Execution Feature Matrix - the category comparison this note joins
References #
https://ollaya.dev/ - product surface: endpoints, 29 model families, platforms, the Ollama 0.40.6 comparison table, Apache-2.0 licensing (re-fetched 2026-10-26)
https://github.com/ollaya-dev/ollaya - repository: Apache-2.0, created 2026-09-23, 1,313 stars as of 2026-10-15, 79 forks, pushed 2026-10-25, releases v0.9.0 through v0.18.0 (GitHub API, as of 2026-10-26)
https://bench.ollaya.dev/ - Ollaya Bench, Edition 2026.10 (2026-10-14, on Ollaya 0.13.0): 36 models over 14 tasks, 5,880 questions per model, RTX 5090 and Threadripper 3960X devices, jeb:27b the leader at 67.8 (fetched 2026-10-17)
https://news.ycombinator.com/item?id=49848269 - the 618-point launch thread (2026-09-25)
https://huggingface.co/ollaya-dev - the pinned model mirrors (15 repositories listed via the Hugging Face API, as of 2026-10-06)