StartLux-Decision v1.1 0.8B Instant

StartLux-Decision v1.1 0.8B Instant is a specialized 0.87B parameter dense decision engine built by StartLux. Unlike standard autoregressive language models that generate responses token by token, this model evaluates decisions in a single forward pass. By framing decisions as typed questions over structured state inputs, it extracts calibrated probability distributions directly across designated outcome tokens. It occupies a distinct space in edge AI and local agent architectures, functioning as an ultra-low-latency classification and routing router rather than a general-purpose conversational assistant.

At 0.87B parameters, it targets local deployments where sub-50 millisecond decision loops are mandatory. Practitioners running agentic pipelines use this model to evaluate environmental states, select discrete tools, adjudicate permission boundaries, and classify programmatic states without the computational overhead or unpredictable latency of multi-token generation. It achieves a score of 53.79 on the public subset of Decision Index 0.3, outperforming standard generative models in its weight class on structured decision benchmarks, though it trades off raw computational and mathematical problem-solving ability to achieve that efficiency.

The model is distributed under a CC BY-NC 4.0 license for research and non-commercial local use with attribution, while commercial deployments require licensing from StartLux Labs. The official inference package and server code are licensed under Apache 2.0, allowing straightforward integration into custom local runtimes via vLLM, llama.cpp, and MLX.

The core architecture is a dense transformer containing 0.87 billion parameters, paired with an expansive 65,536-token context window. While standard generative architectures enforce strict causal masking across the entire sequence, StartLux-Decision v1.1 implements hybrid attention layers. In one out of every four layers, the model applies bidirectional prompt attention ("jev_prompt_attention": "bidirectional" in its configuration), while the remaining linear-attention layers maintain causal properties.

This bidirectional prompt mechanism allows the model to process the provided system state in full view of the target question. During inference, the input prompt supplies the system state (formatted as raw text, structured JSON, or referenced image data) alongside specific decision targets. At the final token position, instead of decoding subsequent text autoregressively, the engine reads the logits corresponding to the candidate options. These logits are divided by a question-specific temperature parameter and normalized using a softmax function over the allowable choices.

Key architectural specifications:

• Parameter Count: 0.87B parameters (873 million active parameters, fully dense)
• Context Length: 65,536 tokens
• Attention Mechanism: Hybrid causal with periodic bidirectional prompt attention (1 in 4 full-attention layers)
• Decision Primitives: Choice (discrete categorical options), Noul (binary yes or no evaluation), and Score (scalar rating on a 2 to 10 scale)
• Forward Pass Execution: Non-generative output extracted directly from candidate logits at prompt termination
• Licensing: CC BY-NC 4.0 for model weights; Apache 2.0 for the inference server and runner

Because the bidirectional attention pattern is non-standard for standard causal text generators, standard inference engines must utilize compatible patches or the dedicated StartLux runtime. Serving the model through a purely causal pipeline alters the attention mask and degrades decision calibration.

StartLux-Decision v1.1 0.8B Instant is designed to sit inside local automation loops, robotics controllers, and programmatic workflow agents. Because it answers in a single pass without token decoding, it provides predictable execution timing for real-time applications.

When building agent frameworks, routing execution based on intermediate tool outputs usually consumes excessive compute when routed through larger models. This model evaluates structured state payloads (such as API responses, system logs, or JSON context objects) and makes precise conditional routing decisions. It reliably handles access permission gates, schema validation branching, and failure recovery state machines.

Using the choice primitive, the model selects appropriate tools from a predefined set based on current conversational or execution state. On benchmarks like the Berkeley Function Calling Leaderboard (BFCL), the architecture registers high accuracy (scoring 93.5 in evaluations) because the bidirectional prompt attention processes parameter requirements alongside the query before emitting the selection probability.

In monitoring pipelines, the noul primitive performs fast binary checks: confirming whether a process completed successfully, verifying if an extracted artifact matches security rules, or checking for anomaly conditions. Processing happens in tens of milliseconds, making it suitable for continuous evaluation loops.

The model is explicitly not suited for tasks that require arithmetic calculations, complex multi-step logical deduction, or free-form text synthesis. When an evaluation requires intermediate computational steps, its single-pass nature causes degradation compared to reasoning-focused models.

Running a local AI model with 0.87B parameters in 2026 demands negligible system overhead compared to typical 7B or 14B architectures. Because the model bypasses iterative token generation, StartLux-Decision v1.1 0.8B Instant performance is measured in prompt evaluation latency per request rather than conventional tokens per second.

Due to the compact 0.87B parameter footprint, memory consumption is primarily determined by context size rather than weight storage. When assessing StartLux-Decision v1.1 0.8B Instant VRAM requirements, keep in mind that pushing context toward the 65,536-token ceiling increases KV cache demands.

The best quantization for StartLux-Decision v1.1 0.8B Instant in production environments is usually BF16 or Q8_0. Because decision calibration relies on precise logit distributions over option letters, aggressive sub-4-bit quantization can shift margin probabilities. If your hardware permits, run unquantized BF16 to maintain optimal scoring calibration.

The minimal StartLux-Decision v1.1 0.8B Instant hardware requirements allow it to run on almost any modern workstation or embedded development board:

• Entry-Level Consumer GPUs: An Nvidia GeForce RTX 3060 (12GB) or RTX 4060 (8GB) can hold the full BF16 model and an extensive context buffer entirely in VRAM. How to run an 0.87B model on consumer GPU setups is straightforward: allocate 2 GB for the model weights, leaving the remainder of the VRAM for high-concurrency request pooling.
• High-End Hardware: On high-tier silicon like the Nvidia RTX 4090, RTX 5090, or professional hardware like the H200, individual decision latency drops below 20 milliseconds per request. While an enterprise accelerator like an H200 delivers the lowest latency, the best GPU for StartLux-Decision v1.1 0.8B Instant from a cost-to-performance standpoint is an entry-level workstation card like an RTX 4060 Ti or RTX 4070.
• Apple Silicon: Unified memory Macs (including baseline M2, M3, or M4 chips with 8GB to 16GB RAM) run the model effortlessly via MLX, offering single-pass latencies under 40 milliseconds.
• CPU and Edge Devices: Because parameter access occurs only once per decision, modern x86 and ARM processors (such as AMD Ryzen 9 series or Raspberry Pi 5 with 8GB RAM using Q4_K_M) achieve practical response times under 150 milliseconds.

To run StartLux-Decision v1.1 0.8B Instant locally, you have multiple engine choices:

1. StartLux Server Engine: The recommended path is the official repository runner (startlux_decide), which exposes an HTTP API compatible with System One /v1/systemone. It correctly configures bidirectional attention masks across the prompt tokens.
2. vLLM Integration: Run via vLLM using the StartLux custom model plugin to enable batching across high-throughput agent environments.
3. llama.cpp / Ollama: Standard GGUF distributions can be executed using llama.cpp builds containing the StartLux attention patch. While Ollama provides an accessible deployment mechanism for local models, ensure the underlying runner supports non-causal prompt attention flags so decision calibration remains intact.

Evaluating StartLux-Decision v1.1 0.8B Instant vs traditional small language models highlights the distinction between generative models and purpose-built decision models.

Qwen2.5-0.5B is a traditional causal language model optimized for generic text generation. When tasked with structured tool calling or categorical classification, Qwen requires structured output engines (such as grammar-constrained decoding) to output valid JSON, generating multiple tokens sequentially. This introduces latencies between 200 and 600 milliseconds on modest hardware. StartLux-Decision resolves the same task in a single forward pass, providing direct logit distributions in under 50 milliseconds. However, Qwen retains the ability to synthesize explanations and write text, which StartLux-Decision cannot do.

Meta's Llama 3.2 1B Instruct offers broader conversational fluency and stronger baseline arithmetic reasoning. When prompts require calculating numeric constraints before making a choice, Llama 3.2 1B performs more reliably because it can emit intermediate reasoning steps. Conversely, on deterministic policy evaluation, permission validation, and high-frequency classification tasks, StartLux-Decision delivers superior calibration (scoring lower Expected Calibration Error) and vastly lower latency, while consuming roughly half the compute footprint during evaluation.