VERIFIED RELEASES DEEPSEEK-R1 & LLAMA-3.3 QUANTIZATION ACTIVE

Foundation Weights.
Optimized For Research.

An open academic registry dedicated to high-throughput distribution of quantized large language models, vision backbones, and scientific checkpoints under permissive licenses.

Explore Models Hub GGUF / SafeTensors Stream Token Demo

248.6 TB

Total Weights Distributed

14.2M

Global Model Pulls

128k

Max Context Horizon

99.98%

Registry Uptime SLA

Open Model Zoo

Verified Model Weights

REASONING Downloads: 4.8M

DeepSeek-R1-Distill Q4_K_M

State-of-the-art open reasoning model distilled from 671B MoE. Superb mathematical derivation and code synthesis capability.

Format: GGUF Arch: MoE / Dense

Download Checkpoint (35MB)

FOUNDATION Downloads: 6.2M

Meta-Llama-3.3-70B Instruct

Industry standard 70-billion parameter transformer. Optimized for long-context dialogue, retrieval-augmented generation and reasoning.

Format: GGUF Context: 128k

Download Checkpoint (50MB)

CODE / SYNTAX Downloads: 2.1M

Qwen-2.5-Coder-32B GGUF

Dedicated coding backbone supporting 92 programming languages. Competitive with proprietary closed-source code engines.

Format: GGUF Pass@1: 86.4%

Download Checkpoint (25MB)

DIFFUSION Downloads: 3.4M

FLUX.1-Schnell FP8

12-billion parameter rectified flow transformer for 4-step photorealistic image generation and spatial rendering.

Format: SafeTensors Latency: 4 Steps

Download Checkpoint (40MB)

ENTERPRISE Downloads: 1.7M

Mistral-Large-2407 123B

Flagship multilingual model with 128k context window. High precision in complex reasoning and function calling.

Format: GGUF / ONNX MMLU: 84.0%

Download Checkpoint (50MB)

AUDIO / ASR Downloads: 5.1M

Whisper-v3-Turbo 809M

Optimized multi-lingual automatic speech recognition model delivering 8x faster transcription with near-zero WER degradation.

Format: PyTorch / ONNX Languages: 99

Download Checkpoint (25MB)

Live WebSockets Engine

Inference Streaming Terminal

Interactive playground proving low-latency token streaming across anycast nodes.

ws://stream.tensorbase.org/v1/chat/completions

Active Model: deepseek-r1

USER QUERY: Optimize an asynchronous parallel matrix multiplication pipeline in Rust with AVX-512 vectorization.

TENSORBASE INFERENCE STREAM (120 TOKENS/SEC): use std::arch::x86_64::*;

// Vectorized matrix multiplication kernel with 512-bit register unrolling
#[inline(always)]
pub unsafe fn matmul_kernel_avx512(a: *const f32, b: *const f32, c: *mut f32) {
    let mut acc = _mm512_setzero_ps();
    for k in 0..16 {
        let va = _mm512_loadu_ps(a.add(k * 16));
        let vb = _mm512_loadu_ps(b.add(k * 16));
        acc = _mm512_fmadd_ps(va, vb, acc);
    }
    _mm512_storeu_ps(c, acc);
}█

Deployment Ecosystem

Native Runtime Compatibility

TensorBase foundation checkpoints integrate seamlessly with production inference frameworks.

ENGINE 01 C/C++

llama.cpp

Bare-metal quantization runtime with zero external dependencies and AVX-512 acceleration.

./llama-cli -m model.gguf

ENGINE 02 HIGH-PPS

vLLM Engine

PagedAttention kernel management delivering 24x serving throughput for concurrent LLMs.

vllm serve tensorbase/r1

ENGINE 03 NVIDIA

TensorRT-LLM

GPU-accelerated enterprise framework utilizing FP8 and INT4 Tensor Cores compilation.

trtllm-build --checkpoint ...

ENGINE 04 DEVELOPER

Ollama CLI

Containerized local model orchestration for rapid testing, prototyping and embeddings.

ollama run tensorbase/llama

Foundation Weights. Optimized For Research.