A Non-Generative System One Architecture for Direct Probabilities
The vLLM Semantic Router (vllm-sr) team has officially released the Decision 3.0 foundation model family under the permissive Apache-2.0 license. Unlike standard autoregressive large language models that generate text token-by-token to explain or categorize inputs, Decision 3.0 operates as a dedicated non-generative System One inference engine designed strictly for high-throughput decision-making.
The models process inputs across text, structured JSON, images (up to 1.6 megapixels), and video clips (sampled at 2 frames per second up to 32 frames). Instead of producing natural language output, a single forward pass yields calibrated probabilities for typed schemas such as multiple Choice, Yes/No determinations, and numerical Score requests without text generation overhead.
Six Parameter Tiers: From On-Device Edge to 26.1B Flagship
Built upon Qwen3.5 and Qwen3.8 foundational architectures, Decision 3.0 spans six parameter footprints: d3-edge (0.59B), d3-lite (0.85B), d3-nano (2.2B), d3-mini (4.5B), d3-flash (8.4B), and the flagship d3 (26.1B / 27B class). These models integrate dedicated vision encoders ranging from 0.10B to 0.46B parameters.
According to self-reported vendor benchmarks on the Jev Decision Index 0.3.1, flagship d3 recorded a score of 64.7 on text tasks—exceeding Perplexity Decider v1.1 (62.8) and TypeSafe Jev (60.1)—and 71.6 on the vision benchmark, surpassing Decider v1.1 (70.6). Median batch-1 latency on a single AMD Instinct MI325X accelerator reached 55 ms for text, 273 ms for images, and 553 ms for 10-second video clips.
Architectural Bottlenecks and Benchmark Verification Nuances
Despite supporting batch question prompts within a unified API call, Decision 3.0 harbors architectural trade-offs. The model does not share intermediate forward passes across concurrent queries; each distinct question triggers its own pass and must reread associated visual context, preventing linear efficiency gains across multiple queries on identical inputs.
Furthermore, benchmark achievements stem from internal vendor runs against published baselines. On selected independent evaluation suites, alternative 27B options such as Torchcast Decision 27B maintain a slight edge, registering 65.1 against 64.9 for d3. Technical teams must treat vendor-published rankings as directional guidance pending independent community audits.
Industry Sentiment: Commoditization Pressures on Proprietary Decision Services
Engineers and machine learning practitioners have responded with intense discussion around market defensibility. The launch of permissive Apache-2.0 models with competitive accuracy puts immediate pricing pressure on proprietary offerings, such as commercial services whose backers raised substantial rounds at premium valuations.
At the same time, enterprise developers highlight operational barriers to adopting self-hosted models. While open weights provide zero-data-retention guarantees, closed SaaS providers already possess formal enterprise compliance clearance inside corporate environments. Deploying internal inference stacks for models like d3 requires dedicated hardware allocation and security reviews, creating an adoption bottleneck despite software availability.
Strategic Takeaways for Thai Enterprises and Cloud Deployments
For technology teams across Thailand, Decision 3.0 provides an actionable mechanism to decouple high-volume routing and compliance pipelines from expensive offshore generative APIs. In workflows such as automated document routing, identity verification screen checks, and e-commerce image moderation, using targeted decision models significantly curtails token costs.
Moreover, smaller tiers like d3-edge and d3-nano run efficiently on modest edge servers or local branch infrastructure, assisting compliance with Thailand's Personal Data Protection Act (PDPA) by processing sensitive visual and textual records fully on-premises without cross-border data transfers. Larger organizations with private cloud hardware can run d3 natively to achieve deterministic classifications within dozens of milliseconds.
Agentic routing, moderation, and structured classification do not require expensive, slow text generation from frontier models. Decision 3.0 brings multimodal single-forward-pass classification to open weights, enabling sub-60ms self-hosted decision pipelines.