---
license: other
language:
- en
library_name: pytorch
pipeline_tag: text-generation
tags:
- text-generation
- custom-llm
- closed-architecture
- untrained
- research-preview
- funding-needed
- smoke-benchmark
- mesko
gated: true
extra_gated_heading: "Mesko Legacy V2 access request"
extra_gated_prompt: "Mesko Legacy V2 is an untrained closed-architecture research preview. Please share how you plan to use, train, evaluate, or fund this project."
extra_gated_fields:
  "Full name": text
  "Organization or lab": text
  "Role": text
  "Country": country
  "Intended use": text
  "Dataset you plan to use": text
  "Dataset license or source": text
  "Training stack or technology": text
  "Compute resources available": text
  "Benchmark plan": text
  "Will share benchmark results with MesklinTech": checkbox
  "Interested in funding or donating compute": checkbox
  "I understand this is an untrained research preview": checkbox
---

# Mesko Legacy V2 LLM

**Mesko Legacy V2 LLM** is a closed-architecture assistant-model research preview by **MesklinTech**. It is published as a gated Hugging Face project to attract training partners, compute sponsors, benchmark contributors, and early reviewers.

## Project Snapshot

| Item | Mesko Legacy V2 |
| --- | --- |
| Release type | Closed-architecture research preview |
| Model state | Untrained V2 concept |
| Access | Manual gated access |
| Goal | Train a proprietary assistant-style LLM from scratch |
| Funding need | GPU compute, dataset preparation, benchmarking, safety testing |
| Architecture detail | Private |

## Short Architecture Line

Custom closed-architecture LLM for scalable assistants.

## What Makes This Project Strong

- **Founder-led and resource-aware:** built from the ground up by a small team focused on training under real compute limits.
- **Closed-model direction:** designed for controlled development similar in spirit to commercial assistant systems.
- **Benchmark-first culture:** even the earliest release includes a smoke benchmark instead of only a vision statement.
- **Gated collaboration:** access requests ask for intended use, dataset plan, training stack, compute, benchmark plan, and funding interest.
- **Ready for serious partners:** the project is structured for compute donors, researchers, labs, and early technical collaborators.

## Smoke Benchmark

This is a smoke test of the training path, not a full model-quality benchmark.

| Metric | Result |
| --- | ---: |
| Previous smoke loss | `1.4551` |
| Improved smoke loss | `0.000249` |
| Target loss gate | `< 0.3000` |
| Train examples | `4` |
| Validation examples | `2` |
| Epochs | `50` |
| Device | CPU |
| Status | Passed |

The smoke test verifies that the from-scratch training path can optimize, checkpoint, validate, and generate on a tiny controlled dataset.

## Comparison With Other LLM Release Styles

| Project style | Example | Public position | How Mesko Legacy V2 differs |
| --- | --- | --- | --- |
| Fully open trained LLM | LLM360 K2 / K2-V2 | Trained weights, data, code, and benchmark reports are released for reproducibility. | Mesko Legacy V2 is closed-architecture and currently seeking funding before full training. |
| Preview LLM release | Trillion-style preview projects | Presents a model direction and early release story before broad adoption. | Mesko Legacy V2 adds gated access and asks collaborators to share dataset, training, and benchmark plans. |
| Untrained architecture preview | Small untrained Hugging Face architecture projects | Shares an early architecture/project idea before real training. | Mesko Legacy V2 keeps architecture private and publishes only high-level status plus smoke-test evidence. |
| Commercial closed model | ChatGPT-style closed assistants | Architecture and training stack are not public, but the product is trained and served at scale. | Mesko Legacy V2 follows a closed-model direction but still needs training compute and funding. |

## Funding Request

We are seeking donations, sponsorship, compute credits, or research collaboration to train Mesko Legacy V2 properly.

Support will be used for:

- GPU training runs
- legally safe dataset preparation
- tokenizer and data pipeline work
- benchmark reporting
- safety testing
- controlled deployment tooling

If you want to help, request access through this gated repo and mention whether you can contribute compute, funding, benchmark work, or training guidance.

## One Clear Limitation

Mesko Legacy V2 is not trained yet, so the current smoke result proves training mechanics only and should not be compared to production benchmark scores.

## Responsible Use

Do not market downstream systems as powered by Mesko Legacy V2 until trained weights, evaluation reports, and usage terms are released.