
KnowLine-31B-Gen2 is a decision model from PelaAI, fine-tuned from google/gemma-4-31B-it. It takes a state (text, a conversation or an image) plus up to 64 typed questions and returns a probability distribution over the answer options in one forward pass, with no text generation. Weights are Apache-2.0 and served with SGLang; a readout temperature of 1.27 is folded into the weights so confidence tracks accuracy. It scores 73.00 on the Decision Index 0.3 public suite, self-run.
A workable 31.27B-parameter dense decision model from PelaAI. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | VRAM |
|---|---|---|
| Acer Veriton GN100 AI MiniAcer | SS | 19.7 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | SS | 19.7 GB |
| AMD Instinct MI300XAMD | SS | 19.7 GB |
| AMD Instinct MI325XAMD | SS | 19.7 GB |
| AMD Instinct MI350PAMD | SS | 19.7 GB |
Cheapest current cloud rentals with at least 20 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA RTX PRO 4000 BlackwellVast.ai · Spot · 24 GB VRAM | $0.11 |
NVIDIA GeForce RTX 3090Vast.ai · Spot · 24 GB VRAM | $0.11 |
NVIDIA RTX A5000RunPod · Community · 24 GB VRAM | $0.16 |
NVIDIA GeForce RTX 3090Vast.ai · On-Demand · 24 GB VRAM | $0.16 |
NVIDIA RTX PRO 5000 BlackwellVast.ai · Spot · 48 GB VRAM | $0.19 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.
KnowLine-31B-Gen2 is a dense 31.27-billion-parameter decision model released by PelaAI under the Apache-2.0 license. Unlike standard autoregressive language models that generate free-form text token by token, KnowLine is built as a dedicated System One decision engine fine-tuned from google/gemma-4-31B-it. It accepts an input state, such as raw text or an ongoing conversation thread, alongside up to 64 structured questions, returning calibrated probability distributions over possible answers in a single forward pass without generating conversational output.
By bypassing autoregressive token generation, KnowLine-31B-Gen2 eliminates the parsing overhead, non-deterministic formatting failures, and multi-token latency penalties common to standard LLMs tasked with classification. The release incorporates a fitted readout temperature of 1.27 folded directly into the model weights, aligning confidence levels directly with predictive accuracy. On the Decision Index 0.3 public evaluation suite, KnowLine-31B-Gen2 achieves a score of 73.00, outperforming sibling architectures and rival 31B-class decision models across agent routing, reasoning, and intent classification benchmarks.
For engineers designing autonomous agents and high-throughput local pipelines, this release fills a specific operational role. Instead of dedicating an entire conversational model to simple branching logic, tool selection, or guardrail validation, KnowLine-31B-Gen2 operates as a low-latency, deterministic routing layer capable of handling dense decision matrices locally.
KnowLine-31B-Gen2 is based on Google's dense Gemma-4-31B-it architecture. PelaAI trained the model using low-rank adaptation (LoRA) targeted specifically at the language model components (rank 32, alpha 64, trained over one epoch on decision datasets curated by teacher models) before merging the adapters back into 16-bit brain floating-point (BF16) weights.
The model relies on Gemma 4's native tokenizer and handles text inputs in English, Simplified Chinese, and Traditional Chinese. Although the underlying base model includes vision components, KnowLine-31B-Gen2's inference script targets text-only decision workloads.
Key architectural characteristics include:
knowline_server.py) implements a read-twice execution strategy for short sequences. By evaluating the input sequence twice through the model's key-value mechanisms, the architecture stabilizes attention weights on complex conditional statements before extracting final logits.KnowLine-31B-Gen2 supports three core decision primitives:
Because the model outputs structured probability arrays instead of strings, developers do not need regex parsers, JSON mode constraints, or grammar-guided decoding layers. Common local workloads include:
Deploying this model locally requires understanding its non-generative nature. While standard LLMs are benchmarked on tokens per second, KnowLine-31B-Gen2 performance is measured in per-request forward-pass latency. With serving through SGLang, a median single-question request takes approximately 0.168 seconds on modern workstation hardware.
The unquantized BF16 checkpoint consumes approximately 62.5 GB of disk space. When loaded via SGLang, the serving engine typically reserves additional headroom for KV cache and scratch space, bringing the recommended baseline to around 70 to 80 GB of video memory.
The official distribution uses SGLang (sglang==0.5.21) behind an API adapter wrapper:
1# Environment setup2pip install "sglang==0.5.21" "transformers==5.12.1" requests34# Launch the backend and decision server5bash serve_knowline.sh PelaAI/KnowLine-31B-Gen2 0 8080
The service exposes an HTTP endpoint (/v1/systemone) accepting payload schemas containing the environment state and an array of typed questions:
1curl -s http://127.0.0.1:8080/v1/systemone -H 'Content-Type: application/json' -d '{2 "model": "PelaAI/KnowLine-31B-Gen2",3 "state": "Customer: The display cracked inside the original packaging before delivery.",4 "questions": {5 "refund_eligibility": {6 "type": "noul",7 "instructions": "Should the agent issue a replacement or refund immediately?"8 },9 "customer_sentiment": {10 "type": "choice",11 "instructions": "Assess customer tone.",12 "criteria": {"frustrated": "Frustrated", "neutral": "Neutral", "satisfied": "Satisfied"}13 }14 }15}'
If you must run KnowLine-31B-Gen2 on consumer hardware with 24 GB of VRAM, 4-bit weight formats are necessary. However, because this model's primary feature is probability calibration (ECE ~0.01), aggressive quantization can disrupt logit margins. If exact probabilities are required to drive safety-critical thresholding, run the model in native BF16 across two 24 GB cards or utilize AWQ/GPTQ formats that preserve output logit distributions better than legacy scalar quantizations.
Evaluating KnowLine-31B-Gen2 against other local AI models requires distinguishing between generative 30B-class models and dedicated decision engines.
| Model | Parameters | DI 0.3 Score | Execution Type | Best Hardware Target |
|---|---|---|---|---|
| KnowLine-31B-Gen2 | 31.27B Dense | 73.00 | Single forward pass (Logits) | 1x 96GB or 2x 24GB (quantized) |
| karpo-decision-31B | ~31.3B Dense | 75.05 | Single forward pass (Logits) | 1x 80GB/96GB GPU |
| KnowLine-26B-A4B-Gen2.5 | ~26B (MoE) | 68.21 | Single forward pass (Logits) | 1x 48GB GPU |
| Gemma-4-31B-it (Base) | 31.27B Dense | N/A | Autoregressive generation | 1x 80GB GPU or 2x 24GB |
Karpo holds a slight lead on raw Decision Index public scores (75.05 vs. 73.00), showing marginally stronger baseline reasoning on abstract edge cases. However, KnowLine-31B-Gen2 exhibits tighter probability calibration on structured tool-routing tasks (scoring 88.8) and benefits from an Apache-2.0 license, making it more practical for self-hosted commercial agent runtimes.
Using an un-tuned generative model like Gemma-4-31B-it for classification requires prompt framing, strict schema constraints (via outlines or JSON mode), and variable multi-token decoding. KnowLine eliminates decoding latency entirely. While generating a response from base Gemma-4 can take 500ms to 2000ms depending on output length, KnowLine delivers evaluated distributions across dozens of questions in less than 200ms on equivalent hardware, operating with complete determinism.

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every PelaAI model we track.

Check which of your devices can run this model and how fast.

Compare hosted per-token prices before you commit to local hardware.

Break-even math for self-hosting, renting a GPU, or paying per token.
Live status for the hosted APIs you might use instead of running this locally.