Releases·Model Release
Dev kgrozdanovski releases CalDec v1, an open-weights decision model for fully-local personal assistants
A developer without professional ML experience released CalDec v1, an open-source decision model with weights, recipe, and dataset, aimed at real-time fully-local personal assistants. Built by fine-tuning open models like Laya and GLiNER on a small dataset, some checkpoints outscored Jev on specific benchmarks, though Jev remains the stronger all-rounder.
2 engagement1 sourceFri, Oct 2, 2026, 01:06 PM
AI Overview
- First-time public model release fine-tunes open-weights models like Laya toward Jev's decision performance using a small dataset
- Some CalDec checkpoints score better than Jev on LocalLLaMA/typed-decisions and the internal test set, though the author says Jev remains the stronger all-rounder
- Releases include CalDec Laya, CalDec GLiNER (gliner2.5-decide), the assistant-decisions dataset on Hugging Face, and code on GitHub
- Target use case is a real-time, fully-local Jarvis-like personal assistant, citing privacy and eliminating network latency
- Author, lacking professional ML experience, ran the experiment to measure how far open-weights models trail Jev
Related
Social·Community
Jev already has rivals: 3 open-source alternatives that are faster, can see, or run locally
A Medium post argues that three weeks after System One models emerged, the category already includes open-source rivals to Jev. The author compares CLM-8B, Clef and Laya on speed, vision capability, and local runnability.
A Medium post argues that three weeks after System One models emerged, the category already includes open-source rivals to Jev. The author compares CLM-8B, Clef and Laya on speed, vision capability, and local runnability.
0 engagement1 sourcerssFri, Oct 9, 2026, 04:06 AM
Social·Community
System 1 Models: Faster and Cheaper Proxy to Frontier Models — With a Working Router
Two Medium posts examine TypeSafe's System One family of decision-focused models — Jev, Laya, Kev, and Nimble — covering their architecture, typed decisions, parallel inference, and calibration. A related post demonstrates routing requests through a LiteLLM proxy to pick the appropriate model. The posts position these open-weights models as faster, cheaper proxies to frontier models for decision workloads rather than conversation.
Two Medium posts examine TypeSafe's System One family of decision-focused models — Jev, Laya, Kev, and Nimble — covering their architecture, typed decisions, parallel inference, and calibration. A related post demonstrates routing requests through a LiteLLM proxy to pick the appropriate model. The posts position these open-weights models as faster, cheaper proxies to frontier models for decision workloads rather than conversation.
0300 engagement3 sourcesrssThu, Sep 24, 2026, 12:00 AM
Social·Community
Benchmarking small confidence scoring decision models: Jev vs Laya evals show label wording and calibration drive accuracy [P]
A practitioner ran independent evals of two small confidence-scoring classification models, TypeSafe AI's hosted Jev and the open-source Laya, for tasks like AML transaction labeling. Results show label descriptions written from input characteristics, uniform time windows, and calibration quality matter more than model choice. Poorly calibrated confidence thresholds can silently cost coverage and accuracy in production filters.
A practitioner ran independent evals of two small confidence-scoring classification models, TypeSafe AI's hosted Jev and the open-source Laya, for tasks like AML transaction labeling. Results show label descriptions written from input characteristics, uniform time windows, and calibration quality matter more than model choice. Poorly calibrated confidence thresholds can silently cost coverage and accuracy in production filters.
1 engagement1 sourceredditWed, Sep 30, 2026, 12:41 PM
Releases·Model Releasefeatured
Developer releases Laya, a 421M-parameter open-source decision model trained with RLCD, with HF space, benchmarks, and repo
Convai Innovations released Laya, an Apache 2.0 non-autoregressive decision model built by Nandakishor M as an open alternative to TypeSafe's closed Jev API. It answers typed questions (choice, score, noul) in a single ~35 ms forward pass, with checkpoints on ModernBERT-large and mmBERT-base, and a community C++ implementation followed within days.
Convai Innovations released Laya, an Apache 2.0 non-autoregressive decision model built by Nandakishor M as an open alternative to TypeSafe's closed Jev API. It answers typed questions (choice, score, noul) in a single ~35 ms forward pass, with checkpoints on ModernBERT-large and mmBERT-base, and a community C++ implementation followed within days.
Updated Sun, Sep 27, 2026, 11:55 PM — Community member lkarlslund released laya.cpp, an MIT-licensed C++ inference implementation with custom CUDA kernels and
211860 engagement4 sourcesredditFri, Sep 18, 2026, 06:25 AM
Releases·Tool Release
Nokia Applied Research releases AnyJev, turning any LLM into a Jev-style typed-decision model with real probabilities and no training
Nokia Applied Research, with a Tencent Hunyuan collaborator, released AnyJev, an open-source Python package (Apache-2.0, on PyPI) that converts any LLM into a Jev-style decision model producing typed decisions and real probabilities with no training. The release demonstrates that raw logit readout is order-sensitive — Qwen3-8B flips a BANKING77 answer with 1.00 confidence when options are reversed — while AnyJev's L0 mode stays consistent with zero labels.
Nokia Applied Research, with a Tencent Hunyuan collaborator, released AnyJev, an open-source Python package (Apache-2.0, on PyPI) that converts any LLM into a Jev-style decision model producing typed decisions and real probabilities with no training. The release demonstrates that raw logit readout is order-sensitive — Qwen3-8B flips a BANKING77 answer with 1.00 confidence when options are reversed — while AnyJev's L0 mode stays consistent with zero labels.
144 engagement1 sourcegithubMon, Sep 21, 2026, 05:55 AM