StartLux-Decision v1.1 27B Instant

StartLux-Decision v1.1 27B Instant is a 27.78B parameter open-weights decision model built specifically for deterministic judgment, classification, and real-time control. Developed by StartLux Labs, the model discards conventional autoregressive text decoding in favor of single-pass probability readouts. When given an environment state along with structured queries, it computes calibrated probabilities across all possible answers in one forward pass, eliminating text generation entirely.

This design targets a distinct category in the open-weights ecosystem: low-latency decision engines for autonomous loops, game playing, policy enforcement, and tool routing. The model scores 72.88 on the public subset of Decision Index 0.3 and executes at a median latency of 52 ms per request on an NVIDIA H200. Licensed under CC-BY-NC-4.0 for non-commercial research, commercial deployments require a separate agreement from StartLux Labs.

Deploying a local AI model 27.78B parameters 2026 release like this presents clear engineering advantages over using standard generative LLMs for classification. Because the architecture extracts terminal logits directly rather than sampling tokens one by one, it eliminates token-budget exhaustion, JSON parsing errors, and decoding overhead.

StartLux-Decision v1.1 27B Instant utilizes a dense transformer architecture containing 27.78B parameters with a native context window of 65,536 tokens. Unlike purely causal autoregressive models, this model employs a hybrid attention layout defined in its configuration as "jev_prompt_attention": "bidirectional".

• Hybrid Attention Layers: In one out of every four layers, the model activates full bidirectional attention across the input prompt, while linear-attention layers remain causal. This enables the model to read earlier sections of the environment state with full visibility into the target question placed at the end of the prompt.
• Logit-Based Extraction: At the final prompt position, the engine reads the logits assigned to the lettered answer choices. These raw logits are scaled by type-specific temperature values and normalized across the valid candidate set.
• Input Envelope: The 65,536 token context window accommodates large structured states (such as serialized JSON trees, game state ledgers, or historical logs) alongside typed questions and up to 8 image inputs.
• Inference Compatibility: The model exposes an interface compatible with the System One /v1/systemone specification. Standard causal runtimes can load the base weights, but running an inference engine that lacks bidirectional prompt support degrades accuracy by altering the intended attention patterns.

Training followed a two-stage regime: an initial full-weight fine-tune across approximately 800,000 decision items, followed by a merged adapter stage trained on 1.7 million additional examples under bidirectional prompt attention.

StartLux-Decision v1.1 27B Instant handles discrete decision-making where deterministic, calibrated probabilities are required. The model processes three core typed question primitives:

• choice: Multi-class selection across discrete, arbitrary lettered options.
• noul: Binary true/false or yes/no verification.
• score: Direct scalar evaluation mapped across a discrete 2 to 10 scale.

Because the system evaluates states without conversational fluff, it serves operational pipelines that require sub-100 ms execution:

• High-Frequency Agent Routing: Acting as a state router, evaluating tool-call safety, or verifying agent plans against rigid permission policies before execution.
• Real-Time Simulation and Robotics: In telemetry monitoring, the model can poll sensor feeds or video frames at fixed intervals (such as 1 Hz) to classify anomalous states, issue flight or hoist commands, or execute real-time game actions in complex environments like StarCraft II or chess.
• Deterministic Guardrails: Verifying user input or database output against compliance criteria, returning a numerical risk confidence without the latency overhead of a generative safety model.

Engineers looking to run StartLux-Decision v1.1 27B Instant locally must plan for the memory footprint of a dense 27.78B parameter model while accounting for its unique execution behavior.

Memory requirements depend on the precision format. Unlike generative LLMs, KV cache growth is minimal because the model executes in a single forward pass per query and does not generate extended output sequences.

• BF16 / FP16 (Unquantized): Requires ~56 GB to 60 GB of VRAM. This configuration mandates an NVIDIA A100 (80GB), H100, H200, or a dual-GPU workstation setup using two RTX 3090 or RTX 4090 cards via pipeline parallelism.
• 8-bit Quantization (GGUF Q8_0 / MLX 8-bit): Requires ~30 GB to 34 GB of VRAM. Fits comfortably on professional enterprise cards like the RTX 6000 Ada, an A5000/A6000 pair, or unified memory Apple Silicon systems (M2/M3/M4 Max with 48GB or 64GB RAM).
• 4-bit Quantization (GGUF Q4_K_M / MLX 4-bit): Requires ~17 GB to 19 GB of VRAM. This is the optimal configuration to run a 27.78B model on consumer GPU hardware such as a single 24GB RTX 3090 or RTX 4090.

The best GPU for StartLux-Decision v1.1 27B Instant in a dedicated single-card workstation is an NVIDIA RTX 4090 (24GB) using 4-bit weights, or an RTX 6000 Ada (48GB) if running at higher precisions. For Mac workstations, an M4 Max with at least 48GB of unified memory running the official MLX implementation yields high local throughput.

The best quantization for StartLux-Decision v1.1 27B Instant is Q4_K_M for 24GB GPUs and 8-bit for systems with 32GB or more headroom.

Because decision models rely on small margin differences between candidate logits at the terminal token position, aggressive low-bit quantizations (such as 2-bit or lower-tier 3-bit variants) can disrupt probability calibration. Q4_K_M preserves logit distribution accuracy on the DecisionPivot benchmark while keeping total VRAM consumption below 20 GB under full 65k context utilization.

When practitioners look for StartLux-Decision v1.1 27B Instant tokens per second metrics, standard benchmarks do not apply directly. Because the model computes answers during the prefill forward pass rather than through autoregressive generation, output throughput is measured in request latency:

• NVIDIA H200 (BF16, vLLM): ~52 ms median per decision request.
• Dual RTX 4090 (FP16, vLLM / custom engine): ~85 to 110 ms per request.
• Single RTX 4090 (Q4_K_M, llama.cpp): ~120 to 160 ms per request.
• Apple M3/M4 Max (MLX 4-bit): ~140 to 180 ms per request.

Official patches and engine support exist across three major local inference frameworks:

1. vLLM (Recommended for Linux/NVIDIA): Provides the lowest latency through batched prefill optimizations and native bidirectional prompt attention support.
2. llama.cpp / Ollama: GGUF weights can be served locally through llama.cpp. When deploying via custom Ollama Modelfiles or llama.cpp builds, apply the official patches/llama.cpp provided in the repository to prevent the engine from falling back to purely causal prompt masking.
3. MLX: Apple Silicon users can run the native 4-bit and 8-bit MLX distributions directly for zero-configuration local execution on macOS.

Evaluating StartLux-Decision v1.1 27B Instant performance against alternative models highlights key functional tradeoffs.

Qwen2.5-32B-Instruct is a general-purpose text generator. When used for decisions, it must ingest a prompt, generate tokens step-by-step (often using structured outputs to ensure valid JSON), and output text.

• Latency: Qwen2.5-32B takes 300 ms to 1,500 ms to generate a response, depending on token count. StartLux-Decision 27B Instant evaluates the query in 52 ms to 150 ms in a single step.
• Calibration: Qwen's logit probabilities are scattered across various syntactic tokens. StartLux explicitly computes normalized, calibrated probabilities over target options.
• Flexibility: Qwen can provide explanatory justifications and perform freeform coding. StartLux Instant does not generate explanations; it strictly outputs decision probabilities.

Within the same family, the AutoThink variant introduces conditional reasoning.

• Behavior: The Instant model forces every decision through a single forward pass. AutoThink evaluates a confidence gate (0.6 by default). If the highest probability is below this threshold, AutoThink pauses, generates a compact reasoning trace, and conducts a second readout pass.
• Tradeoff: Instant is strictly deterministic in latency, making it the right pick for real-time robotic loops, game ticks, or systems with rigid SLA caps. AutoThink achieves higher score accuracy (73.37 vs 72.88 on Decision Index 0.3) at the expense of variable, generation-dependent latency on complex or ambiguous edge cases.