UraionLabs

uraion-forge-2b

README

License: apache-2.0

Overview

Uraion Forge 2B is an ultra-compact, high-performance language model engineered specifically for edge execution, multi-turn agentic tool calling, and high-throughput quantitative reasoning. Developed by Uraion Labs, Forge 2B transforms a 2-billion parameter foundation into a deterministic, production-grade sub-agent.

Built on the robust MiniCPM5-2B architecture (Llama-style transformer with Grouped-Query Attention and 131k RoPE context), Forge 2B features targeted post-training regularizations that resolve the most notorious failure modes of small reasoning models: token-budget exhaustion ("runaway thinking"), parameter hallucination, and multi-turn state drift.

Whether deployed across enterprise production clusters via PyTorch / Transformers / vLLM, or on edge devices, Uraion Forge 2B delivers enterprise-tier agentic reliability with deterministic execution.

Available Formats & Ecosystem

Table with columns: Package, Format / Framework, Primary Use Case, Target Hardware, Repository Link
Package	Format / Framework	Primary Use Case	Target Hardware	Repository Link
Base Model (This Repo)	Safetensors (BF16 / FP16)	Universal deployment, PyTorch, Transformers, vLLM, SGLang, TGI	NVIDIA GPUs (CUDA), AMD (ROCm), Linux, Windows, macOS	****
Apple Silicon Native	MLX (4-bit, 8-bit, 16-bit)	Bare-metal Mac inference (90-120+ tok/s), unified memory	Apple Silicon (M1/M2/M3/M4/M5/M6)	****
GGUF & Dynamic Quants	GGUF (Q2 to F16)	Ollama, llama.cpp, LM Studio, mobile / low-VRAM edge	CPU, Metal, Vulkan, Edge Devices	****

Key Model Highlights

⚡ Blazingly Fast Local Edge Execution: At 2B active parameters, Forge 2B fits comfortably within 4 GB to 11 GB of unified memory, executing at over 90+ tokens/second on Apple Silicon M-series chips via native MLX and modern NVIDIA RTX GPUs via PyTorch.
🧠 Concise Chain-of-Thought (CoT) Regularization: Eliminates the classic small-model pathology of runaway reasoning loops. Forge 2B is trained to allocate internal <think>...</think> tokens proportionally to task complexity, or bypass internal CoT entirely for instant, direct programmatic answers.
🛠️ Autonomous Agentic Tool Reliability (90.5%): Outperforms comparable compact models on multi-turn API workflows. Native support for optimistic concurrency resolution (e.g. CONFLICT $\to$ re-read $\to$ re-acquire locks), idempotent retry handling, and complex multi-parameter JSON schema adherence.
❓ Active Disambiguation & Clarification: Replaces arbitrary parameter hallucination with explicit clarification queries. When critical function parameters are omitted, Forge 2B halts execution and prompts the user with an interrogative question ending in '?'.
📈 Zero Degradation Capability Retention: Retains 100% of core pre-trained general knowledge while achieving 75.0% long-context needle retrieval (+16.7 pp improvement over base) and .

Technical Specifications

Table with columns: Parameter, Specification, Notes
Parameter	Specification	Notes
Base Architecture	`MiniCPM5-2B` (`LlamaForCausalLM`)	Pinned commit `abe115e887989b14f05e64a3b260648329324c3f`
Active Parameters	2,048,286,720 (2.05B)	Full model active during forward pass
Layers	42	Transformer decoder layers
Hidden Dimension	2048	Model representation width

Quickstart Guides

1. Apple Silicon MLX Quickstart (`mlx-lm`)

Uraion Forge 2B is natively tuned for Apple Silicon Unified Memory via Apple's MLX framework:

bash
pip install mlx-lm

python
from mlx_lm import load, generate

# 1. Load Forge 2B
model, tokenizer = load("uraionlabs/uraion-forge-2b")

# 2. Prepare structured chat prompt
messages = [
    {
        "role": "system",
        "content": "You are Uraion Forge, an advanced autonomous reasoning and coding assistant developed by Uraion Labs."
    },
    {
        "role": "user",
        "content": "Write an efficient Python class implementing an asynchronous ring buffer with thread-safe append and pop methods."
    }
]

prompt_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# 3. Generate response with optimal temperature
response = generate(
    model=model,
    tokenizer=tokenizer,
    prompt=prompt_text,
    max_tokens=1024,
    temp=0.2,
    verbose=True
)
print(response)

2. PyTorch & Hugging Face Transformers Quickstart

Deploy on NVIDIA GPUs or CPU instances using standard Hugging Face Transformers:

bash
pip install transformers torch accelerate

python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("uraionlabs/uraion-forge-2b", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "uraionlabs/uraion-forge-2b",
    torch_dtype=dtype,
    device_map="auto" if device == "cuda" else None,
    trust_remote_code=True
)
if device != "cuda":
    model.to(device)

# Formulate prompt
messages = [
    {
        "role": "system",
        "content": "You are Uraion Forge, an advanced autonomous reasoning and coding assistant developed by Uraion Labs."
    },
    {
        "role": "user",
        "content": "Write an optimized Python algorithm for computing rolling volume-weighted average price (VWAP) over streaming ticks."
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=1024,
        temperature=0.2,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)

3. Structured Agentic Tool Calling & Disambiguation

Uraion Forge 2B supports deterministic JSON function calling and zero-hallucination clarification loops:

python
import json

# Define your tools schema
tools = [
    {
        "type": "function",
        "function": {
            "name": "query_timeseries_metric",
            "description": "Query historical metric time series data for an asset or system component.",
            "parameters": {
                "type": "object",
                "properties": {
                    "asset_id": {"type": "string", "description": "Asset or cluster symbol."},
                    "metric_name": {"type": "string", "description": "Metric name (e.g., 'volume_vwap', 'latency_p99')."},
                    "window_minutes": {"type": "integer", "description": "Window duration (1 to 1440 minutes)."}
                },
                "required": ["asset_id", "metric_name"]
            }
        }
    }
]

system_prompt = (
    "You are Uraion Forge, an autonomous agent capable of utilizing external tools and APIs.\n"
    "When invoking a tool, respond with a JSON markdown code block matching the specified schema.\n"
    "If a required parameter is omitted by the user, DO NOT guess or hallucinate the parameter; "
    "instead, formulate a concise, direct clarification query terminating with a question mark ('?').\n\n"
    f"Available Tools:\n{json.dumps(tools, indent=2)}"
)

# Example 1: Fully-specified query -> Generates structured JSON tool call
user_msg_1 = "Retrieve the 60-minute volume_vwap metric for asset 'URAI-ALPHA'."
# Forge 2B Output:
# ```json
# {
#   "name": "query_timeseries_metric",
#   "arguments": {
#     "asset_id": "URAI-ALPHA",
#     "metric_name": "volume_vwap",
#     "window_minutes": 60
#   }
# }
# ```

# Example 2: Missing required parameter -> Generates clarification query ending in '?'
user_msg_2 = "Query the timeseries metric for asset 'URAI-ALPHA'."
# Forge 2B Output:
# "Which metric_name (such as 'volume_vwap' or 'latency_p99') would you like to query for asset 'URAI-ALPHA'?"

Empirical Benchmark Performance

Uraion Forge 2B was rigorously evaluated across unseen confirmation benchmarks, development suites, and general capability retention probes.

1. Sealed Confirmation Benchmark Performance

Evaluated on 41 sealed, held-out tasks (20 programmatic coding specifications and 21 multi-turn agentic workflows across 6 distinct software environments):

Table with columns: Mode / Configuration, Task Domain, Pass Rate, Successful / Total, Truncated Turns, Generation Latency
Mode / Configuration	Task Domain	Pass Rate	Successful / Total	Truncated Turns	Generation Latency
Agentic Tool Confirmation (Primary)	Tool Workflows	90.5%	19 / 21	2 / 21	2,043s
Agentic Tool Confirmation (Secondary)	Tool Workflows	61.9%	13 / 21	0 / 21	273s

[!TIP] Production Recommendation: For direct code synthesis and programmatic calculations, operate the model in Secondary Mode (thinking: false). This reduces truncation errors by 87.5% (from 16 down to 2) and increases execution throughput tenfold. For complex multi-turn API workflows requiring state tracking, enable Primary Mode (thinking: true) for 90.5% tool reliability.

2. General Capability Retention & Long-Context Suite

To verify that specialized agentic training did not degrade foundation capabilities, Forge 2B was audited across the 252-probe General Retention Suite against the untouched base foundation:

Table with columns: Evaluation Domain, Base Model, Uraion Forge 2B, Delta, Status
Evaluation Domain	Base Model	Uraion Forge 2B	Delta	Status
Long-Context Needle Retrieval (16k–32k)	58.3% (7/12)	75.0% (9/12)	+16.7 pp	🚀 Substantial Gain
Multiple Choice Reasoning (MMLU subset)	60.1% (137/228)	58.8% (134/228)	-1.3 pp	✅ Preserved
Structured JSON Schema Extraction	91.7% (11/12)

3. Numerical Verification & Export Drift

Every release artifact is structurally audited across 381 tensor weights:

FP32 Reference Merge: Bit-level numerical equivalence verified with mean KL divergence of $1.55 \times 1 0 - 7$ and 100.0% top-1 token agreement against the unfused dynamic LoRA adapter.
Serving Smoke Pass: 100% verified pass rate on local loopback HTTP serving (http://127.0.0.1:8899/v1) across structured JSON output, sandboxed Python code execution, and multi-turn native tool calls.

High-Throughput Quantitative Reasoning & Safeguards

Uraion Forge 2B is trained to excel at general quantitative, mathematical, and algorithmic reasoning:

Data Structures & Streaming: High-performance ring buffers, lock-free queues, exponential moving averages, and streaming statistics.
Concurrency & State Management: Multi-turn optimistic concurrency controls, version conflict detection (409 Conflict), and asynchronous job polling.
Strict Privacy Guarantees: Uraion Forge 2B was post-trained exclusively on public synthetic code, standard open benchmarks, and structural API interfaces. No proprietary trading alpha, execution strategies, order flow mechanisms, or internal financial datasets are present in this model.

Citation & Contact

If you use Uraion Forge 2B in your research or applications, please cite:

bibtex
@misc{uraionlabs2026forge2b,
  title={Uraion Forge 2B: High-Throughput Quantitative Reasoning and Autonomous Agentic Tool Inference at the Edge},
  author={{Uraion Labs Technical Team}},
  year={2026},
  howpublished={\url{https://huggingface.co/uraionlabs/uraion-forge-2b}},
  note={Uraion Labs Open Source Release}
}

For inquiries, enterprise deployments, and partnership information, visit Uraion Labs.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

UraionLabs

Model Tree

Base

openbmb/MiniCPM5-2B

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Overview

Available Formats & Ecosystem

Table with columns: Package, Format / Framework, Primary Use Case, Target Hardware, Repository Link
Package	Format / Framework	Primary Use Case	Target Hardware	Repository Link
Base Model (This Repo)	Safetensors (BF16 / FP16)	Universal deployment, PyTorch, Transformers, vLLM, SGLang, TGI	NVIDIA GPUs (CUDA), AMD (ROCm), Linux, Windows, macOS	****
Apple Silicon Native	MLX (4-bit, 8-bit, 16-bit)	Bare-metal Mac inference (90-120+ tok/s), unified memory	Apple Silicon (M1/M2/M3/M4/M5/M6)	****
GGUF & Dynamic Quants	GGUF (Q2 to F16)	Ollama, llama.cpp, LM Studio, mobile / low-VRAM edge	CPU, Metal, Vulkan, Edge Devices	****

Key Model Highlights

⚡ Blazingly Fast Local Edge Execution: At 2B active parameters, Forge 2B fits comfortably within 4 GB to 11 GB of unified memory, executing at over 90+ tokens/second on Apple Silicon M-series chips via native MLX and modern NVIDIA RTX GPUs via PyTorch.
🧠 Concise Chain-of-Thought (CoT) Regularization: Eliminates the classic small-model pathology of runaway reasoning loops. Forge 2B is trained to allocate internal <think>...</think> tokens proportionally to task complexity, or bypass internal CoT entirely for instant, direct programmatic answers.
🛠️ Autonomous Agentic Tool Reliability (90.5%): Outperforms comparable compact models on multi-turn API workflows. Native support for optimistic concurrency resolution (e.g. CONFLICT $\to$ re-read $\to$ re-acquire locks), idempotent retry handling, and complex multi-parameter JSON schema adherence.
❓ Active Disambiguation & Clarification: Replaces arbitrary parameter hallucination with explicit clarification queries. When critical function parameters are omitted, Forge 2B halts execution and prompts the user with an interrogative question ending in '?'.
📈 Zero Degradation Capability Retention: Retains 100% of core pre-trained general knowledge while achieving 75.0% long-context needle retrieval (+16.7 pp improvement over base) and .

Technical Specifications

Table with columns: Parameter, Specification, Notes
Parameter	Specification	Notes
Base Architecture	`MiniCPM5-2B` (`LlamaForCausalLM`)	Pinned commit `abe115e887989b14f05e64a3b260648329324c3f`
Active Parameters	2,048,286,720 (2.05B)	Full model active during forward pass
Layers	42	Transformer decoder layers
Hidden Dimension	2048	Model representation width

Quickstart Guides

1. Apple Silicon MLX Quickstart (`mlx-lm`)

Uraion Forge 2B is natively tuned for Apple Silicon Unified Memory via Apple's MLX framework:

bash
pip install mlx-lm

python
from mlx_lm import load, generate

# 1. Load Forge 2B
model, tokenizer = load("uraionlabs/uraion-forge-2b")

# 2. Prepare structured chat prompt
messages = [
    {
        "role": "system",
        "content": "You are Uraion Forge, an advanced autonomous reasoning and coding assistant developed by Uraion Labs."
    },
    {
        "role": "user",
        "content": "Write an efficient Python class implementing an asynchronous ring buffer with thread-safe append and pop methods."
    }
]

prompt_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# 3. Generate response with optimal temperature
response = generate(
    model=model,
    tokenizer=tokenizer,
    prompt=prompt_text,
    max_tokens=1024,
    temp=0.2,
    verbose=True
)
print(response)

2. PyTorch & Hugging Face Transformers Quickstart

Deploy on NVIDIA GPUs or CPU instances using standard Hugging Face Transformers:

bash
pip install transformers torch accelerate

python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("uraionlabs/uraion-forge-2b", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "uraionlabs/uraion-forge-2b",
    torch_dtype=dtype,
    device_map="auto" if device == "cuda" else None,
    trust_remote_code=True
)
if device != "cuda":
    model.to(device)

# Formulate prompt
messages = [
    {
        "role": "system",
        "content": "You are Uraion Forge, an advanced autonomous reasoning and coding assistant developed by Uraion Labs."
    },
    {
        "role": "user",
        "content": "Write an optimized Python algorithm for computing rolling volume-weighted average price (VWAP) over streaming ticks."
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=1024,
        temperature=0.2,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True)
print(response)

3. Structured Agentic Tool Calling & Disambiguation

Uraion Forge 2B supports deterministic JSON function calling and zero-hallucination clarification loops:

python
import json

# Define your tools schema
tools = [
    {
        "type": "function",
        "function": {
            "name": "query_timeseries_metric",
            "description": "Query historical metric time series data for an asset or system component.",
            "parameters": {
                "type": "object",
                "properties": {
                    "asset_id": {"type": "string", "description": "Asset or cluster symbol."},
                    "metric_name": {"type": "string", "description": "Metric name (e.g., 'volume_vwap', 'latency_p99')."},
                    "window_minutes": {"type": "integer", "description": "Window duration (1 to 1440 minutes)."}
                },
                "required": ["asset_id", "metric_name"]
            }
        }
    }
]

system_prompt = (
    "You are Uraion Forge, an autonomous agent capable of utilizing external tools and APIs.\n"
    "When invoking a tool, respond with a JSON markdown code block matching the specified schema.\n"
    "If a required parameter is omitted by the user, DO NOT guess or hallucinate the parameter; "
    "instead, formulate a concise, direct clarification query terminating with a question mark ('?').\n\n"
    f"Available Tools:\n{json.dumps(tools, indent=2)}"
)

# Example 1: Fully-specified query -> Generates structured JSON tool call
user_msg_1 = "Retrieve the 60-minute volume_vwap metric for asset 'URAI-ALPHA'."
# Forge 2B Output:
# ```json
# {
#   "name": "query_timeseries_metric",
#   "arguments": {
#     "asset_id": "URAI-ALPHA",
#     "metric_name": "volume_vwap",
#     "window_minutes": 60
#   }
# }
# ```

# Example 2: Missing required parameter -> Generates clarification query ending in '?'
user_msg_2 = "Query the timeseries metric for asset 'URAI-ALPHA'."
# Forge 2B Output:
# "Which metric_name (such as 'volume_vwap' or 'latency_p99') would you like to query for asset 'URAI-ALPHA'?"

Empirical Benchmark Performance

Uraion Forge 2B was rigorously evaluated across unseen confirmation benchmarks, development suites, and general capability retention probes.

1. Sealed Confirmation Benchmark Performance

Evaluated on 41 sealed, held-out tasks (20 programmatic coding specifications and 21 multi-turn agentic workflows across 6 distinct software environments):

Table with columns: Mode / Configuration, Task Domain, Pass Rate, Successful / Total, Truncated Turns, Generation Latency
Mode / Configuration	Task Domain	Pass Rate	Successful / Total	Truncated Turns	Generation Latency
Agentic Tool Confirmation (Primary)	Tool Workflows	90.5%	19 / 21	2 / 21	2,043s
Agentic Tool Confirmation (Secondary)	Tool Workflows	61.9%	13 / 21	0 / 21	273s

[!TIP] Production Recommendation: For direct code synthesis and programmatic calculations, operate the model in Secondary Mode (thinking: false). This reduces truncation errors by 87.5% (from 16 down to 2) and increases execution throughput tenfold. For complex multi-turn API workflows requiring state tracking, enable Primary Mode (thinking: true) for 90.5% tool reliability.

2. General Capability Retention & Long-Context Suite

To verify that specialized agentic training did not degrade foundation capabilities, Forge 2B was audited across the 252-probe General Retention Suite against the untouched base foundation:

Table with columns: Evaluation Domain, Base Model, Uraion Forge 2B, Delta, Status
Evaluation Domain	Base Model	Uraion Forge 2B	Delta	Status
Long-Context Needle Retrieval (16k–32k)	58.3% (7/12)	75.0% (9/12)	+16.7 pp	🚀 Substantial Gain
Multiple Choice Reasoning (MMLU subset)	60.1% (137/228)	58.8% (134/228)	-1.3 pp	✅ Preserved
Structured JSON Schema Extraction	91.7% (11/12)

3. Numerical Verification & Export Drift

Every release artifact is structurally audited across 381 tensor weights:

FP32 Reference Merge: Bit-level numerical equivalence verified with mean KL divergence of $1.55 \times 1 0 - 7$ and 100.0% top-1 token agreement against the unfused dynamic LoRA adapter.
Serving Smoke Pass: 100% verified pass rate on local loopback HTTP serving (http://127.0.0.1:8899/v1) across structured JSON output, sandboxed Python code execution, and multi-turn native tool calls.

High-Throughput Quantitative Reasoning & Safeguards

Uraion Forge 2B is trained to excel at general quantitative, mathematical, and algorithmic reasoning:

Data Structures & Streaming: High-performance ring buffers, lock-free queues, exponential moving averages, and streaming statistics.
Concurrency & State Management: Multi-turn optimistic concurrency controls, version conflict detection (409 Conflict), and asynchronous job polling.
Strict Privacy Guarantees: Uraion Forge 2B was post-trained exclusively on public synthetic code, standard open benchmarks, and structural API interfaces. No proprietary trading alpha, execution strategies, order flow mechanisms, or internal financial datasets are present in this model.

Citation & Contact

If you use Uraion Forge 2B in your research or applications, please cite:

bibtex
@misc{uraionlabs2026forge2b,
  title={Uraion Forge 2B: High-Throughput Quantitative Reasoning and Autonomous Agentic Tool Inference at the Edge},
  author={{Uraion Labs Technical Team}},
  year={2026},
  howpublished={\url{https://huggingface.co/uraionlabs/uraion-forge-2b}},
  note={Uraion Labs Open Source Release}
}

For inquiries, enterprise deployments, and partnership information, visit Uraion Labs.

uraion-forge-2b

README

Overview

Available Formats & Ecosystem

Key Model Highlights

Technical Specifications

Quickstart Guides

1. Apple Silicon MLX Quickstart (mlx-lm)

2. PyTorch & Hugging Face Transformers Quickstart

3. Structured Agentic Tool Calling & Disambiguation

Empirical Benchmark Performance

1. Sealed Confirmation Benchmark Performance

2. General Capability Retention & Long-Context Suite

3. Numerical Verification & Export Drift

High-Throughput Quantitative Reasoning & Safeguards

Citation & Contact

Explore FriendliAI today

README

Overview

Available Formats & Ecosystem

Key Model Highlights

Technical Specifications

Quickstart Guides

1. Apple Silicon MLX Quickstart (mlx-lm)

2. PyTorch & Hugging Face Transformers Quickstart

3. Structured Agentic Tool Calling & Disambiguation

Empirical Benchmark Performance

1. Sealed Confirmation Benchmark Performance

2. General Capability Retention & Long-Context Suite

3. Numerical Verification & Export Drift

High-Throughput Quantitative Reasoning & Safeguards

Citation & Contact

1. Apple Silicon MLX Quickstart (`mlx-lm`)

1. Apple Silicon MLX Quickstart (`mlx-lm`)