vLLM Semantic Router releases Decision-2.0-Lux-9B, a Qwen3.5-9B fine-tune answering six decision questions in one pass

Releases·Model Release

Shanghai AI Lab releases Intern-Decision-4B and 0.8B multimodal structured decision models

InternLM released Intern-Decision-4B, a multimodal structured decision model fine-tuned from Qwen3.5-4B, alongside a 0.8B variant, with model weights, a demo, and a GitHub repo. The models take a shared state, a question schema, and optional images, and return answer distributions for all questions in a single forward pass without free-form generation. This gives practitioners a deterministic structured-candidate-scoring API for decision tasks rather than autoregressive text sampling.

InternLM released Intern-Decision-4B, a multimodal structured decision model fine-tuned from Qwen3.5-4B, alongside a 0.8B variant, with model weights, a demo, and a GitHub repo. The models take a shared state, a question schema, and optional images, and return answer distributions for all questions in a single forward pass without free-form generation. This gives practitioners a deterministic structured-candidate-scoring API for decision tasks rather than autoregressive text sampling.

Releases·Model Releasefeatured

TypeSafe AI launches Jev, an RLCD-trained decision model claiming 20-200x faster, 40-400x cheaper inference than LLMs

TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, released Jev on September 15, 2026 after two years in stealth — a frontier model that takes unstructured state plus typed questions and returns parallel, calibrated decisions instead of generated text. It is trained with a new objective, Reinforcement Learning for Calibrated Decisions (RLCD), positioned against RLHF and RLVR. TypeSafe claims roughly two orders of magnitude faster and cheaper inference than LLMs on decision tasks, and community reaction has been intense: follow-up analyses, a JevBench benchmark, and open-weight imitators like Kev appeared within days.

TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, released Jev on September 15, 2026 after two years in stealth — a frontier model that takes unstructured state plus typed questions and returns parallel, calibrated decisions instead of generated text. It is trained with a new objective, Reinforcement Learning for Calibrated Decisions (RLCD), positioned against RLHF and RLVR. TypeSafe claims roughly two orders of magnitude faster and cheaper inference than LLMs on decision tasks, and community reaction has been intense: follow-up analyses, a JevBench benchmark, and open-weight imitators like Kev appeared within days.

Releases·Model Release

vLLM SR releases Decision 3.0: Apache-2.0 multimodal decision models 0.6B–27B, #1 on both Jev Decision Index boards

The vLLM Semantic Router team released Decision 3.0 (d3), an Apache-2.0 family of decision models from 0.6B to 27B built on Qwen3.5 and Qwen3.8-27B. The models output probabilities for choice, yes/no, and score questions without text generation, reading text, images, and video. The 27B ranks #1 on both Jev Decision Index boards, and every size leads its class.

The vLLM Semantic Router team released Decision 3.0 (d3), an Apache-2.0 family of decision models from 0.6B to 27B built on Qwen3.5 and Qwen3.8-27B. The models output probabilities for choice, yes/no, and score questions without text generation, reading text, images, and video. The 27B ranks #1 on both Jev Decision Index boards, and every size leads its class.

Dragonfly is a LoRA fine-tune of Qwen3-VL that reads an image once and scores many questions in the same pass, outputting calibrated probabilities instead of generated text. It solves the cost of asking a VLM one question per pass. Built by an undergrad in Finland as a weekend side project.

Researchers introduce SeLMRoute, an LLM routing framework that decouples interpretable semantic evidence extraction from performance estimation and deployment objectives. A decision model assesses queries via probability-distributed judgments, and a lightweight supervised router uses this semantic state to pick models. It outperforms the strongest fixed candidate on LLMRouterBench and supports both performance- and cost-aware routing objectives.