StartLux-Decision v1.1 2B Instant

StartLux-Decision v1.1 2B Instant is a specialized open-weights decision model developed by StartLux Labs. Built as a 2.27B parameter dense architecture, the model departs from standard autoregressive text generation. Instead of decoding tokens sequentially to construct an answer, it ingests an environment state alongside typed questions and outputs calibrated probabilities for every available option in a single forward pass. This single-pass design delivers deterministic, high-throughput classification and policy selection without the latency or parsing errors common to generative LLMs.

Released under a CC BY-NC 4.0 license with an Apache-2.0 inference runtime, this 2.27B variant is engineered for practitioners who need fast, local routing, structured grading, or discrete decision loops. On the public subset of Decision Index 0.3, the model posts a score of 59.23, outperforming general-purpose small models prompted for classification. By skipping text generation entirely, it offers sub-50ms decision loops on local hardware, making it a viable component for autonomous agent triage, real-time game bots, and safety filter pipelines.

For developers deploying models on their own hardware, StartLux-Decision v1.1 2B Instant provides a rare balance of high decision quality and minimal compute requirements. With its compact footprint, you can run this local AI model with 2.27B parameters in 2026 on modest consumer graphics cards or unified memory laptops, offloading decision logic from expensive cloud APIs to local edge devices.

The model utilizes a dense transformer architecture totaling 2.27B parameters. Unlike standard decoder-only models that rely exclusively on causal attention masking, StartLux-Decision implements a hybrid attention scheme. One out of every four layers features full bidirectional attention across the prompt, designated in its configuration as "jev_prompt_attention": "bidirectional". The remaining three layers retain causal linear attention. This hybrid design allows the model to read the entire input state with the question already in view, conditioning early representations on the target options before extracting probabilities at the final position.

Training was conducted in two primary stages. Stage 1 involved full-parameter fine-tuning over approximately 2.7 million structured decision examples. Stage 2 trained an adapter with bidirectional prompt attention across another 1.7 million decision items, which was subsequently merged into the base weights.

Inference does not rely on a standard generation loop. A user prompt presents a context state (formatted as text or JSON, supporting up to 8 image inputs) followed by a question with lettered options. At the final sequence position, the raw logits corresponding to the option identifiers are extracted, divided by a task-specific temperature stored in the configuration, and normalized over the valid option subset via softmax. Because context length is consumed purely as prompt context for a single forward evaluation, the model avoids the memory churn and compute overhead of maintaining an expanding KV cache during decoding.

StartLux-Decision v1.1 2B Instant is strictly calibrated for discrete evaluations rather than creative prose or coding generation. It accepts up to 26 options per read and supports three explicit question schemas:

• Choice (`choice`): Multi-class classification where the model picks between listed categorical options (A through Z).
• Binary (`noul`): Direct yes or no determinations, outputting calibrated probability distributions for boolean checks.
• Score (`score`): Ordinal grading on a discrete scale from 2 to 10, mapping states to calibrated intensity, urgency, or quality levels.

Because it yields true probabilities rather than text tokens, downstream software can gate decisions on strict confidence thresholds. If confidence falls below an acceptable delta, your system can defer to human intervention or invoke a larger reasoning model.

Common local engineering use cases include:

• Fast Agent Routing: Dispatching user requests to appropriate specialized tools, workflows, or sub-agents without waiting for a 7B or 8B model to generate JSON.
• Content Moderation and Gating: Running deterministic safety, policy, or permission checks against internal documents and user inputs at line-rate speed.
• Real-Time Game and Simulation Control: Evaluating game states (such as chess boards or RTS unit queues) every few milliseconds to select optimal immediate actions.
• Sensor and Log Triage: Scanning incoming JSON log entries or telemetry states and classifying operational health on a 2 to 10 scale.

Because this architecture relies on single-pass forward evaluation with bidirectional attention, execution behaves differently than standard generative models. Standard implementations run through vLLM, llama.cpp (with the StartLux attention patch), and native MLX on macOS. For local deployments, evaluating StartLux-Decision v1.1 2B Instant tokens per second is less relevant than measuring total requests per second, as every query completes in a single prompt evaluation step without token decoding overhead.

The small parameter count makes StartLux-Decision v1.1 2B Instant hardware requirements exceptionally light. VRAM demands scale based on the chosen quantization:

• BF16 / FP16 (Full Precision): Requires ~4.8 GB VRAM. Recommended for maximum calibration accuracy on cards like the RTX 3060 12GB or RTX 4060 Ti 16GB.
• Q8_0 / 8-bit MLX: Requires ~2.6 GB VRAM. Preserves near-lossless calibration metrics while fitting into ultra-low-power devices.
• Q4_K_M / 4-bit MLX: Requires ~1.5 GB to 1.8 GB VRAM. This is the best quantization for StartLux-Decision v1.1 2B Instant when deploying alongside other models on memory-constrained devices.

• Nvidia Consumer GPUs: An entry-level RTX 3060 (12GB) or RTX 4060 (8GB) can run the model in native BF16 with batch processing room to spare. On an enthusiast card like the RTX 4090, single-request latency drops to single-digit milliseconds. The best GPU for StartLux-Decision v1.1 2B Instant in production clusters is an RTX 4000 Ada or RTX 4090, which easily yields hundreds of decisions per second across parallel streams.
• Apple Silicon: On an M2, M3, or M4 Mac (including base models with 8GB or 16GB unified memory), the official MLX 4-bit and 8-bit ports execute instantly via the unified memory pool without draining thermal headroom.
• CPU and Edge Devices: When quantized to Q4_K_M, the model requires less than 2 GB of system RAM, enabling deployment on mini PCs, edge nodes, or Raspberry Pi 5 clusters.

To run StartLux-Decision v1.1 2B Instant locally, you can use the official StartLux inference engine, which provides a local HTTP endpoint compatible with the System One /v1/systemone specification. If you are integrating into local stacks using Ollama or raw llama.cpp, ensure your backend binary includes the required custom bidirectional prompt attention patch (jev_prompt_attention). Running the weights through an unpatched causal-only engine will cause the model to read the state without attending forward to the question options, degrading calibration.

To evaluate StartLux-Decision v1.1 2B Instant performance, consider how it stacks up against small general-purpose models applied to classification tasks.

• StartLux-Decision v1.1 2B Instant vs Llama 3.2 3B (Instruct): When tasked with classification, Llama 3.2 3B must generate text tokens, requiring parsing logic and risking syntax failure. Prompting Llama for log probabilities on options requires logit extraction hooks that are cumbersome to manage. StartLux-Decision v1.1 2B Instant operates faster, eliminates output parsing, and provides properly normalized calibration out of the box, scoring 59.23 on Decision Index 0.3. However, Llama 3.2 3B remains necessary if you need free-form explanatory text or conversational replies.
• StartLux-Decision v1.1 2B Instant vs Gemma 2 2B: Gemma 2 2B is a capable generative base, but fine-tuning it to output calibrated scores across dynamic option counts (up to 26 choices) requires substantial dataset engineering. StartLux-Decision already handles categorical choices, binary assertions, and scaled grading natively in a single pass.
• StartLux-Decision v1.1 2B Instant vs StartLux-Decision AutoThink Variants: Within its own model family, the Instant model runs strictly in one forward pass. The AutoThink variants (available at 4B, 9B, and 27B) check confidence against a threshold gate (defaulting to 0.6). When unsure, AutoThink generates a brief internal reasoning trace before deciding. Choose the 2B Instant model when low latency, zero token generation, and predictable execution speed take priority over multi-step chain-of-thought analysis.