
GS1-4B is an open-source decision model from BIGAI (Beijing Institute for General Artificial Intelligence). It returns native probability distributions over runtime-defined candidates for agent routing, verification, extraction and control, using Choice, Noul and Score primitives with no generated output tokens. The 4.7B model is a full fine-tune of Qwen3.5-4B with a hybrid Gated DeltaNet and segmented full-attention backbone, supports branch lengths up to 65,536 tokens, and scores 68.97 on the Decision Index 0.4.1 public index. Weights and code are Apache-2.0.
A solid 4.7B-parameter dense decision model from Beijing Institute for General Artificial Intelligence. 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 |
|---|---|---|
| ACEMAGIC M1A Pro (i9-13900HK + ARC A770)ACEMAGIC | SS | 5.7 GB |
| Acer Veriton GN100 AI MiniAcer | SS | 5.7 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | SS | 5.7 GB |
| AMD Instinct MI300XAMD | SS | 5.7 GB |
| AMD Instinct MI325XAMD | SS | 5.7 GB |
Cheapest current cloud rentals with at least 6 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA GeForce RTX 5060 TiVast.ai · Spot · 16 GB VRAM | $0.06 |
NVIDIA Tesla V100 16GBVast.ai · Spot · 16 GB VRAM | $0.08 |
NVIDIA RTX A4000Vast.ai · Spot · 16 GB VRAM | $0.08 |
NVIDIA RTX A4000Vast.ai · On-Demand · 16 GB VRAM | $0.09 |
NVIDIA GeForce RTX 4070Vast.ai · Spot · 12 GB VRAM | $0.09 |
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.
GS1-4B is an open-source, native decision model developed by the Beijing Institute for General Artificial Intelligence (BIGAI). Built as a full fine-tune of Qwen3.5-4B, the model diverges sharply from standard autoregressive text generators. Instead of emitting sequential text tokens to answer prompts, GS1-4B computes native, calibrated probability distributions over runtime-defined candidate sets. It acts as a dedicated System One engine for agentic workflows, handling tasks like tool selection, agent routing, verification, and real-time state control with zero generated output tokens.
With 4.7B parameters (4,704,520,192 stored parameters) and an Apache-2.0 license, GS1-4B addresses a critical bottleneck in local AI agent pipelines: the latency, parsing fragility, and compute overhead of using generative language models for deterministic control flow. On the Decision Index 0.4.1 public benchmark, GS1-4B achieves a score of 68.97 across 43 multi-domain text suites, placing it directly on the parameter-performance Pareto frontier for small-footprint decision backbones.
For local deployments, GS1-4B functions as a low-latency controller that runs comfortably on modern consumer graphics cards. Its specialized architecture allows developers to replace brittle prompt-based JSON classification pipelines with direct, mathematically calibrated probability scores.
GS1-4B is built on a hybrid recurrent-attention backbone derived from Qwen3.5, combining sub-quadratic sequence mixing with targeted attention:
Because the model does not autoregressively sample tokens, inference consists of a single forward pass over the state and candidate set. When using CUDA Graphs, GS1-4B hits a P50 latency of 35.17 ms for single queries on modern accelerators.
GS1-4B operates exclusively through three typed decision primitives:
p(true) probability for any given proposition against the supplied context.In production environments, these primitives enable distinct agentic capabilities:
Noul primitive to verify retrieved chunks or generated statements against reference documents across its 65k context window.Evaluating the hardware requirements for GS1-4B differs from standard text models. Because the model outputs native distributions rather than text, traditional "GS1-4B tokens per second" generation metrics do not apply. Instead, GS1-4B performance is measured in decisions per second and forward-pass latency.
The memory footprint depends on the context length utilized within your agent pipeline:
To run GS1-4B locally, you will need PyTorch 2.9+, Transformers 5.3.0+, and the custom modeling files supplied in the official repository. Because of the custom modular_gs1.py architecture and cross-attention decision head, standard Ollama or basic llama.cpp runtimes do not currently support GS1-4B out of the box. Deployment is best handled via Hugging Face Transformers or BIGAI's native inference runtime using Python 3.11+.
Understanding where GS1-4B fits requires comparing it against standard small language models and traditional classification backbones.
Practitioners often use Qwen2.5-3B or 7B Instruct to route agent queries by constraining generation with outlines, regex, or JSON schemas.
DeBERTa-v3-Large (435M) has long been the standard local choice for zero-shot classification and NLI verification.
For developers architecting agent loops, GS1-4B functions as a fast, deterministic arbiter, offloading control flow from expensive generation models to a dedicated, locally hosted decision engine.

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Beijing Institute for General Artificial Intelligence 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.