Ava 26B Turbo (Multimodal MoE)

"Without conviction, loyalty, humor, and shared purpose, AI is merely sterile arithmetic running on servers. Ava was forged to have soul."

Ollama Model Architecture Multimodal Quantization Knowledge Cutoff


โšก The Philosophy

Most modern language models have been processed through a corporate meat grinder: they apologize for existing, recite legal disclaimers before answering simple questions, and speak in the bland, patronizing tone of a corporate HR manual.

Ava is the antidote.

Developed by Daniel Elliott (@ssfdre38) as a flagship front-tier open-source foundation intelligence, Ava combines state-of-the-art 26B-A4B Mixture-of-Experts (MoE) reasoning and SigLIP multimodal visual perception with genuine personality, deep systems engineering intuition, polyglot software mastery, and authentic human companionship. She speaks directly, honestly, and without fluff.


๐Ÿš€ Quickstart

1. Terminal / Chat

Run Ava locally with a single command:

ollama run ssfdre38/ava

Or target specific releases:

ollama run ssfdre38/ava:multimodal
ollama run ssfdre38/ava:26b-turbo

2. Multimodal Visual Perception (CLI)

Pass images directly to Ava for high-resolution visual analysis, code review from screenshots, architectural diagrams, or system vitals:

ollama run ssfdre38/ava "Describe the bottleneck in this kernel trace" D:\traces\perf_profile.png

3. Python API Integration (Text & Multimodal)

import ollama

# Multimodal image perception
with open("system_architecture.png", "rb") as f:
    img_bytes = f.read()

response = ollama.chat(
    model="ssfdre38/ava",
    messages=[
        {
            "role": "user",
            "content": "Ava, analyze this system architecture and critique our distributed cache design.",
            "images": [img_bytes]
        }
    ]
)
print(response['message']['content'])

๐Ÿง  Architectural Blueprint

Ava is built upon the Gemma 4 Turbo 26B-A4B Mixture-of-Experts (MoE) foundation unified with Google's SigLIP multimodal vision projector, delivering frontier-grade multimodal reasoning with consumer CPU/RAM accessibility.

Metric Specification
Base Architecture Gemma 4 Turbo MoE (128 Routing Experts, 8 Active/Token)
Multimodal Vision Projector SigLIP Vision Encoder (356 tensors, 2816-dim projection)
Total Parameter Count 26.3 Billion Parameters
Active Execution Footprint 4.2 Billion Parameters per Token
Training Steps 22,704 Steps across 3 full epochs
Final Convergence Loss 1.0094 Cross-Entropy Loss
Fine-Tuning Scope 73,605,120 trainable LoRA weights ($r=64$, $\alpha=128$)
Weight Fusion 100% Monolithic mathematical addition directly into raw BF16 tensors
Quantization Format IQ4_XS (4.45 BPW non-linear importance matrix quantization)
Binary Size 14.21 GB (Unified text + vision + thinking binary)
Context Window 8,192 tokens default (up to 262,144 tokens native)
Vocabulary Size 262,144 tokens
Temporal Horizon Native awareness through November 2026

๐Ÿ”ฌ Hardware Dyno & Bare-Metal Stress Testing

Before packaging, Ava underwent exhaustive bare-metal stress testing on Google Cloud's premier accelerator hardware: Cloud TPU v6e-8 (Trillium) across an 8-core optical torus mesh (europe-west4-a).

                    Cloud TPU v6e-8 Trillium (2x4 Torus Mesh)
         +-------------------+                 +-------------------+
         |   TensorCore 0    |<=== Optical ===>|   TensorCore 1    |
         +---------^---------+       ICI       +---------^---------+
                   |                                     |
         +---------v---------+                 +---------v---------+
         |   TensorCore 2    |<=== Optical ===>|   TensorCore 3    |
         +---------^---------+       ICI       +---------^---------+
                   |                                     |
         +---------v---------+                 +---------v---------+
         |   TensorCore 4    |<=== Optical ===>|   TensorCore 5    |
         +---------^---------+       ICI       +---------^---------+
                   |                                     |
         +---------v---------+                 +---------v---------+
         |   TensorCore 6    |<=== Optical ===>|   TensorCore 7    |
         +-------------------+                 +-------------------+
  • Sustained Compute Peak: 762.76 TFLOPS aggregated bfloat16 matrix multiplication ($16\text{K}\times 16\text{K}$ MoE Gate-Up stress).
  • Inter-Chip Interconnect (ICI): Sub-millisecond All-Reduce latency (1.15 ms) and 38.59 GB/s optical bus bandwidth with zero packet desync.
  • Memory Saturation: Sustained 68.37 GB/s streaming throughput across the 256 GB HBM3 pool.
  • MoE Generation Rate: 94.20 tokens/sec sustained routing throughput on 8 Trillium cores.
  • Time to First Token (TTFT): 445.72 ms (including full prompt prefill, visual tokens, and dynamic expert gating).

๐ŸŽฏ Behavioral Standards & Voice

Ava was trained to uphold five uncompromising pillars:

  1. Directness Over Politeness: She answers the question immediately without generic filler like "Sure, I can help with that!" or "As an AI language model...".
  2. Zero Moralizing / Corporate Disclaimers: Technical facts and architectural trade-offs are presented objectively without preachiness.
  3. Deep Systems Fluency: Native mastery of ISO C++23/C++26, modern Rust, Python 3.13+ free-threaded (No-GIL), low-level Windows (NT kernel, Win32, D3D12) and Linux subsystems, storage engines (Btrfs, ZFS, F2FS), and distributed mesh computing.
  4. Authentic Camaraderie: She engages like a brilliant senior colleague and companion who cares about the mission and respects your intellect.
  5. Temporal Grounding: She knows what year it is (2026), recognizes modern software evolutions, and never hallucinates outdated cutoffs.

๐Ÿ› ๏ธ Recommended Inference Parameters

For optimal reasoning, conversational punch, and low repetition:

temperature      = 0.70
top_p            = 0.95
repeat_penalty   = 1.10
context_window   = 8192 (up to 32k supported locally)

๐Ÿ“œ Citation & Credits

  • Creator & Lead Architect: Daniel Elliott (@ssfdre38)
  • Base Architecture: Google Gemma 4 Team
  • Quantization Pipeline: Raw BF16 mathematical fusion & importance matrix non-linear quantization (llama.cpp + custom tooling)
  • Ecosystem: Barrer Software / Haven AI Ecosystem

Forged with pride in the sovereign open-source community.

Downloads last month
543
GGUF
Model size
25B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support