Yvligo

Yvligo is an open-weights model announced on 9 November 2026 by the startup log(log(ai)), also called loglog. It is a supervised fine-tune of DeepSeek V4.1 Flash (1017) that acts as a tool-using agent: given a client brief, it sets up, monitors, and adjusts ad campaigns on X (Twitter). The weights are released on Hugging Face as a PEFT LoRA adapter under the MIT license. Co-founder Ted posted the launch note as @notmiramurati.

The same release introduces AdVal, the team's evaluation and simulation of an X ads market. Agents get only the official X Ads MCP (which the simulator serves) and a brief, then have to run the campaign. The model card reports that on 45 held-out briefs Yvligo never saw, its geometric mean of conversions ÷ a scripted baseline (par) is 1.311×, ahead of untuned DeepSeek V4.1 Flash (1.095), Grok 4.7 (1.087), and Gemini 4 Argon (1.198) in that test. The homepage uses a different normalization (share of the best-known conversions on each brief) and lists Yvligo at 77%, versus 70% for Gemini 4 Argon and 64% for Grok 4.7. The launch post also cites a live flight: on a $200 budget, Yvligo is reported to have driven 383 relevant website visits, against 123 for Grok. Those figures are the company's; they are not our measurements.

Highlights:

• Site: https://loglog.ai/ ; blog (dated 9 November 2026): https://loglog.ai/blog/introducing-adval ;
• Weights: https://huggingface.co/loglogai/yvligo ; benchmark repo: https://github.com/loglogai/AdVal ;
• Training: expert-agent campaign traces in the simulator, then LoRA supervised fine-tuning on Tinker (rank 32, the step-39 checkpoint, about 24 million tokens);
• At inference, the card says the model should act through mcp__ads__* tools and finish with a recap to the client.

Why it matters

Evaluating ad-buying agents on a live platform is slow and expensive. This release pairs a reproducible simulator-and-benchmark (AdVal) with an open adapter trained in that environment. The model card publishes geometric means, paired comparisons, and tool-error rates, and the traces are public with the benchmark — more inspectable than a one-line claim of beating Grok by 3×.

If you want to try an agent workflow on X ads, the concrete path is: read AdVal's tasks and traces, attach the LoRA to DeepSeek V4.1 Flash as the card describes, and run your own briefs through the same Ads MCP tools. The team also reports one live $200 comparison. Keep the simulator leaderboard and that single live flight separate. The blog calls real A/B tests "a sobering experience" and says the leaderboard will change as the simulation improves.

Limits

AdVal is the company's own simulation and benchmark; the ranking and the 1.311× figure have not been independently reproduced here. The model card itself lists limits: some of the gain over the base model comes from keyword targeting, which the simulator may reward more than live X does, and in 18 of 45 held-out runs the model's reasoning remarked that it was being evaluated. Rule-following on age gates, devices, and geos is imperfect. The live comparison is a single $200 test scored on "relevant visits," too small to treat as a stable production result. Serving the adapter also requires hosting the DeepSeek V4.1 Flash base model. Check X ads policy, data handling, and local advertising rules before running it on a real account.

1. Read https://loglog.ai/ and https://loglog.ai/blog/introducing-adval , and use the launch post https://x.com/notmiramurati/status/2119872200622940560 to see where the live-test numbers were stated;
2. Get the LoRA from https://huggingface.co/loglogai/yvligo and load it on deepseek-ai/DeepSeek-V4.1-Flash-1017 as the model card shows (Transformers + PEFT);
3. Follow https://github.com/loglogai/AdVal for tasks and traces, and run the card's system prompt, client brief, and mcp__ads__* tool surface — this is not a plain chat adapter;
4. If you move to a real X ads account, start with a small budget and re-check targeting, bids, and policy rejections; do not treat the simulator board or the single $200 test as a budget forecast.