MISA AI 2.0 Decision Plus · oak 1.1

MISA AI 2.0 Decision Plus (codename oak 1.1) is a specialized 31.27B dense decision model developed by MISA JSC. Fine-tuned from Google's Gemma 4 31B base, the model breaks away from standard autoregressive text generation. Instead of sampling tokens sequentially to draft an answer, oak 1.1 operates as an ultra-fast classification and scoring engine. It ingests an input state (such as unstructured text, system logs, dialogue histories, or JSON structures) alongside explicit candidate options, executing a single forward prefill pass to return calibrated probability distributions directly across designated target labels.

For engineers building multi-agent workflows, safety barriers, and triage pipelines, running generative large language models for pure classification is computationally inefficient. Autoregressive decoding introduces latency jitter, formatting inconsistencies, and high compute overhead. By restricting output strictly to soft-capped logits over target options, oak 1.1 functions as a high-throughput deterministic router. It scores 68.19 on the Decision Index public suite, outperforming predecessor iterations and offering sub-100 millisecond decision latency on enterprise hardware.

As a local AI model with 31.27B parameters, oak 1.1 delivers deterministic classification without relying on external cloud APIs. The model weights are released under a Creative Commons Attribution-NonCommercial 4.0 license (CC-BY-NC-4.0), making it freely accessible for private evaluation, academic benchmarking, and self-hosted non-commercial pipelines, while commercial production deployments require commercial licensing agreements from MISA JSC.

The underlying architecture relies on a dense transformer network comprising 31.27 billion parameters across 60 transformer layers with a hidden dimension of 5376 and a vocabulary size of 262,144 tokens. While based on the Gemma4ForConditionalGeneration blueprint, oak 1.1 alters inference execution in two structural ways:

• Bidirectional Full Attention: Unlike standard causal decoders that mask future tokens during prompt evaluation, oak 1.1 enables bidirectional attention across prompt tokens within its global full-attention layers (while preserving native localized sliding-window layers). This noncausal prompt processing ensures the model reads question instructions and label candidates simultaneously with the input context, maximizing semantic cohesion before calculating final probabilities.
• Zero-Token Sampling Logit Head: The inference engine extracts logits at the terminal sequence position, computing a softmax distribution strictly over the token IDs corresponding to valid option keys (such as choice letters). Because generation terminates at prompt prefill, the model incurs no iterative token-generation overhead.
• 256,000 Token Context Window: The architecture supports contexts up to 256K tokens, enabling practitioners to evaluate massive inputs such as financial ledgers, dense technical manuals, repository-level traces, or legal transcripts within a single evaluation pass.

MISA AI 2.0 Decision Plus addresses deterministic operational tasks across English and Vietnamese environments. Every decision request accepts an input state and maps to three supported evaluation primitives:

• `noul` (Binary Verification): Evaluates truth probability for a single premise against the input state. This serves automated fact verification, contract clause validation, and policy compliance checks.
• `choice` (Categorical Selection): Evaluates mutually exclusive multi-class problems. The model returns normalized probabilities for every defined option alongside overall decision confidence. Standard implementations include semantic routing between agent tools, support ticket prioritization, and multi-label intent classification.
• `score` (Ordinal Regression): Projects ordinal rankings over an ordered series of discrete categories (such as urgency levels ranging from low to catastrophic).

To prevent out-of-distribution hallucinations, oak 1.1 returns a metric called label_mass. This metric measures the aggregate raw probability captured strictly by valid candidate labels. If label_mass drops substantially, it indicates that the provided state does not match the decision criteria, alerting the host orchestrator to escalate the request to human review or fallback systems.

Deploying this model locally requires understanding the difference between generative decoders and decision classifiers. When developers ask about MISA AI 2.0 Decision Plus · oak 1.1 tokens per second, the practical metric is actually prefill time and latency per decision. Because it emits no sequential output tokens, latency is defined almost entirely by how fast your system handles prompt ingest. On an NVIDIA H200, oak 1.1 delivers a median latency of 67 ms per request.

The full BF16 checkpoint is approximately 62.6 GB across sharded safetensors. The table below details local hardware allocations based on context depth and weight formats:

• Best GPU for MISA AI 2.0 Decision Plus · oak 1.1: For production clusters, a single 80GB card (NVIDIA A100 SXM4 or H100/H200 NVL) running BF16 ensures baseline latency under 100 ms. For local developer workstations, dual NVIDIA RTX 3090 or RTX 4090 (48GB combined via PCIe) using 8-bit precision offers the ideal balance of accuracy, cost, and high-speed prefill.
• How to run 31.27B model on consumer GPU hardware: If you are restricted to a single 24GB consumer graphics card (such as an RTX 4090 or RTX 3090 Ti), choose the best quantization for MISA AI 2.0 Decision Plus · oak 1.1: Q4_K_M. This quantization keeps the parameter footprint near 18.5 GB, leaving roughly 4 GB to 5 GB of VRAM for activation layers under moderate prompt lengths (4,000 to 8,000 tokens). On Apple Silicon, unified memory architectures (such as an M3 Max or M4 Max with 64GB or 128GB RAM) can easily fit the unquantized or Q8 models with minimal memory pressure.

Standard causal runtimes like basic vLLM or standard Ollama endpoints will process prompts linearly. While compatible endpoints like SGLang serve /v1/systemone out of the box causally, achieving the official benchmark scores requires noncausal bidirectional evaluation. Practitioners should serve the weights using the official linne_engine.py wrapper equipped with nc_hook.py, which injects bidirectional full-attention masks directly into the attention layers:

1import sys
2from huggingface_hub import snapshot_download
3from linne_engine import LinneEngine
4
5# Download model repository
6path = snapshot_download("PNPM244/oak1.1")
7sys.path.insert(0, path)
8
9# Initialize noncausal execution on local GPU
10engine = LinneEngine(model=path, device="cuda:0", noncausal=True)
11
12# Execute single-pass evaluation
13decision, raw_metrics = engine(
14    state="Database replication lag exceeded 1500ms on secondary node db-02.",
15    questions={
16        "severity": {
17            "type": "choice",
18            "instructions": "Determine incident severity triage level.",
19            "criteria": {
20                "P1": "Critical outage affecting customer write path",
21                "P2": "Elevated performance degradation with fallback operating",
22                "P3": "Minor operational anomaly without customer impact"
23            }
24        }
25    }
26)
27
28print(decision["answers"]["severity"])

Evaluating MISA AI 2.0 Decision Plus · oak 1.1 vs alternative open-weight architectures highlights stark operational differences:

• oak 1.1 vs Qwen 2.5 32B Instruct: Qwen 2.5 32B is a general-purpose autoregressive model capable of conversational reasoning, code generation, and complex synthetic thought. However, when deployed purely to classify text or route tool executions, Qwen must generate textual responses, requiring speculative decoding or guided JSON grammars (via Outlines or LMQL). This introduces generation latency (often 500 ms to 2 seconds). oak 1.1 finishes the exact same evaluation in a single forward pass under 100 ms with superior native calibration on structured label sets.
• oak 1.1 vs Gemma 2 27B / Llama 3.3 70B Classifiers: Custom sequence classification heads attached to Llama or Gemma bases typically lose zero-shot versatility when target classes change, requiring dedicated parameter fine-tuning for new label definitions. oak 1.1 processes arbitrary ad-hoc criteria directly inside the prompt context while retaining the pre-trained semantic depth of its 31.27B base.

For developers seeking a specialized, low-latency arbiter to govern LLM agent swarms, enforce content safety boundaries, or triage massive document pipelines locally, oak 1.1 provides an exceptionally fast, deterministic inference profile that generic autoregressive models cannot match.