H2O.ai released H2O-Lightning-4B, an open-weight Apache-2.0 model built for 'decision API' inference: it returns calibrated probabilities from a single forward pass with no generated tokens, making decisions fast and cheap. The 4B model (fine-tuned from Qwen3.5-4B) scores 72.5 on the JevBench public leaderboard, above Jev 1.13 at 71.5, making it the top open model. Announced by an H2O.ai employee with weights on Hugging Face; 12B and 31B versions are planned.
A developer released Mica v0.1 4B, an Apache-2.0 decision model for agent gates and routers that returns calibrated probabilities instead of generating text. Built on Qwen3.5-4B with a rank-16 LoRA and trained on 77,732 rows for under $30 of rented RTX 3090 time. It scores below the closed Jev 1.13 API on benchmarks but resists in-state prompt injection far better (69% vs 18%) and runs in about 47 ms per decision on an RTX 3090.
A developer released Mica v0.1 4B, an Apache-2.0 decision model for agent gates and routers that returns calibrated probabilities instead of generating text. Built on Qwen3.5-4B with a rank-16 LoRA and trained on 77,732 rows for under $30 of rented RTX 3090 time. It scores below the closed Jev 1.13 API on benchmarks but resists in-state prompt injection far better (69% vs 18%) and runs in about 47 ms per decision on an RTX 3090.
EmpirioLabs released Aplomb 1, a 5.3B open-weights decision model built on Qwen3.5-4B that reads up to 1M tokens of text, JSON, images, video and audio in a single request. It returns probabilities for every tool and each enum/boolean argument, letting agents act on confident calls and defer uncertain ones to larger models. It ranks #1 on the Hugging Face Typed Decisions benchmark and #1 among 4B models on Decision Index 0.2.1.
EmpirioLabs released Aplomb 1, a 5.3B open-weights decision model built on Qwen3.5-4B that reads up to 1M tokens of text, JSON, images, video and audio in a single request. It returns probabilities for every tool and each enum/boolean argument, letting agents act on confident calls and defer uncertain ones to larger models. It ranks #1 on the Hugging Face Typed Decisions benchmark and #1 among 4B models on Decision Index 0.2.1.
A business consultant fine-tuned Qwen3.5-4B into ImaJev-4b, a model that makes business decisions from text and up to two photos in one forward pass. It ranked #1 of 91 on JevBench (67.37) and #3 of 56 on DecisionBench, ahead of GPT-5.6 Luna and DeepSeek V4.1. Weights are released under Apache-2.0 with code and a demo, and the total training cost was about $1,200 in rented GPU time.
A business consultant fine-tuned Qwen3.5-4B into ImaJev-4b, a model that makes business decisions from text and up to two photos in one forward pass. It ranked #1 of 91 on JevBench (67.37) and #3 of 56 on DecisionBench, ahead of GPT-5.6 Luna and DeepSeek V4.1. Weights are released under Apache-2.0 with code and a demo, and the total training cost was about $1,200 in rented GPU time.
Unsloth announced an open-source repo and notebooks that turn small LLMs (Qwen3.5 0.8B/4B, Gemma 4, Laya) into locally trainable 'Decision Models' that output decisions instead of text. Training runs on consumer GPUs from 2.5GB to 8GB VRAM, with a UI, guide, and video tutorial. The method lifted Qwen3.5 0.8B aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks.
Unsloth announced an open-source repo and notebooks that turn small LLMs (Qwen3.5 0.8B/4B, Gemma 4, Laya) into locally trainable 'Decision Models' that output decisions instead of text. Training runs on consumer GPUs from 2.5GB to 8GB VRAM, with a UI, guide, and video tutorial. The method lifted Qwen3.5 0.8B aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks.