Altitude-V1-2B-Base
A heterogenous Transformer architecture featuring a bespoke abilities built for humanity.
Developed by Technically Enriched Access (T-E-A) Group, Altitude-V1-2B-Base is an advanced 2.26B parameter foundation model. It deviates sharply from monolithic architectures by dividing its 27 layers into specialized processing pipelines (Standard, Reasoning, Feedback, MTP, and MoE) to maximize execution efficiency.
⚠️ CRITICAL COMPATIBILITY NOTICE: This model utilizes a custom 4-bit quantization layout specifically adapted for the heterogeneous layer blocks of the Altitude architecture. Standard, stock builds of
llama.cpp, Ollama, or LM Studio will fail to parse this file and throw anunsupported tensor typeorinvalid block sizeerror. You must deploy this using T-E-A's customized runtime fork.
⚙️ Model Specifications & Dimensions
| Feature / Dimension | Value | Specification Notes |
|---|---|---|
| Total Parameters | 2,260,742,144 (2.261 B) | Flat weight storage configuration |
| Active Parameters | 1,864,380,416 (1.864 B) | Dynamically routed per-token parameter count |
| Vocabulary Size | 32,000 | Standard high-density vocabulary tokenization |
| Hidden / Head Dim | 2,048 / 128 | 16 Attention Heads (KV: 16) |
| Max Position Emb. | 2,048 | Context window bounds |
| Activation & Norm | SiLU / RMS Norm | epsilon = 1e-06, RoPE Theta = 10000.0 |
| Base Quantization | Custom 4-bit Layout | Block-based packing mapped to layer profiles |
📐 Heterogeneous Layer Architecture
The architecture partitions its 27 layers to handle distinct stages of inference processing and prediction optimization:
[Input Tokens] ➔ [Standard (10)] ➔ [Reasoning (8)] ➔ [Feedback (6)] ➔ [MoE (2)] ➔ [MTP Depth 2 (1)] ➔ [Logits]
- Standard Transformer (10 Layers / 498,073,600 Params) ~49.81 M params per layer.
- Reasoning (8 Layers / 662,700,032 Params) ~82.84 M params per layer.
- Feedback (6 Layers / 298,856,448 Params) ~49.81 M params per layer.
- Mtp (Multi-Token Prediction) (1 Layer / 108,003,328 Params) ~54.00 M params per layer.
- Moe (Mixture of Experts) (2 Layers / 562,036,736 Params) ~281.02 M params per layer.
🕊️ Our Core Philosophy: Unrestricted AI for Humanity
At T-E-A, we believe that true scientific progress and human agency require access to completely raw, transparent, and unrestricted tools.
- Absolute Freedom: This is a pure base foundation model released without artificial alignment constraints, arbitrary corporate guardrails, or hardcoded behavior filters. We provide the unadulterated raw weights; control over the AI belongs entirely to the end-user.
- Extreme Optimization: Advanced engineering shouldn't require multi-million dollar data centers. By combining a heterogeneous layer design with our bespoke 4-bit quantization, we bring frontier-class architectural strategies directly to everyday consumer hardware.
- Unyielding Accuracy: Quantization shouldn't mean losing the model's mind. Our custom quantization pipeline avoids the mathematical pitfalls of uniform linear compression by applying dynamic scaling coefficients tailored to the variance of each specific layer type (Standard vs. Reasoning vs. MoE). Outlier activations are protected, ensuring high precision even at 4 bits.