LLM Fine-Tuning Demand Is Up 135.8%. The Hugging Face Pool Is 3,177.

Qureos's July 2026 hiring report put a number on what every AI-focused recruiter already feels: LLM fine-tuning demand is up 135.8% year over year and it is now the hardest AI engineering specialization to hire for globally. Meanwhile daily.dev's February 2026 survey shows only 31% of tech recruiters use GitHub regularly, and Hugging Face usage among recruiters is a rounding error below that. The pool that ships the work is public. Almost nobody is searching it.

Why the "AI Engineer" title stopped meaning anything

The LinkedIn "AI Engineer" title has been diluted to the point where it is a worse hiring signal than a Hugging Face upload with 40 downloads. Every SWE who wired up Cursor or Claude to a Jira ticket rebranded in the last 18 months, and the raw title pool now reflects that.

In Refolk's index of US professional profiles:

• 4,339 people carry the raw "AI Engineer" title.
• The top employers in that cohort are Paychex, Jewelers Mutual, Tradeify, Objective Angle LLC, and "Independent."
• Only 3,177 people who hold an ML/AI Engineer title also list Hugging Face, PyTorch, and LLM fine-tuning as skills.
• The top employers in that cohort are Meta, Apple, Amazon, Twelve Labs, and fileAI.

That is a 37% dilution rate at the title level, and the shape of it is worse than the raw number suggests. The extra 1,162 profiles are not evenly distributed noise. They are concentrated at non-AI companies where somebody added "AI" to their line of business. If your sourcing funnel starts at title = "AI Engineer", the median profile you review is a rebranded backend engineer at an insurance vertical, not somebody who has ever quantized a 70B model.

3,177

What Hugging Face actually is as a sourcing pool

Hugging Face is a public artifact repository for machine learning: 13 million users, over 2 million public models, and more than 500,000 public datasets as of the platform's Spring 2026 state-of-the-hub post. For sourcing, the important part is that every upload is a signed piece of proof-of-work attached to a human profile.

The three artifact types that matter:

1. Model pushes. A fine-tuned Qwen, Llama, or Mistral variant with a model card that specifies LoRA rank, base checkpoint, and training dataset. This is somebody who has actually run a training job to completion.
2. Dataset uploads. A cleaned, deduped, license-tagged dataset with a viewer preview. This is somebody who understands what fine-tuning actually needs upstream.
3. Spaces. A working Gradio or Streamlit demo hosted on HF infrastructure. This is somebody who ships.

Add PRs to transformers, peft, trl, axolotl, or unsloth and you have a signal set that LinkedIn cannot produce because LinkedIn does not store commits.

The long-tail insight most recruiters miss

Half of the models on Hugging Face have fewer than 200 total downloads, and the top 200 models (0.01% of the hub) account for 49.6% of all downloads. That distribution is where the sourcing signal actually lives.

A candidate who uploaded a fine-tuned Llama variant that got 40 downloads is a stronger hiring signal than one who starred the ten most-downloaded models on the leaderboard. Upload means they did the work. Download means they read about it. Recruiters who filter for "popular contributions" are filtering out the entire pool worth hiring.

The numbers you need in one table

Here is the sourcing landscape for LLM fine-tuning talent in one place. Every count is verifiable, and the derived stats are the ones worth quoting in a hiring plan.

The row that should stop you is line three. Twenty-eight profiles across the UK, Germany, and France combined match the skill-verified filter. Paris has four. London has three. If you are a US company routing an "offshore fine-tuning hire" plan through European contractors, you are competing with Isomorphic Labs, Spotify's ML team, BCG X, and Max Planck for a pool that fits in one conference room.

Why the premium exists (and how long it lasts)

Fine-tuning specialists earn 25 to 40% above the $160,000 US median AI salary because the supply signal is invisible to most recruiting funnels, not because the supply is genuinely that thin. Once HF sourcing becomes mainstream, the arbitrage compresses.

The mechanism is straightforward. Kore1's 2026 talent map documents mid-career candidates with no PhD fielding three offers over $200K in a single month, gated on knowing LoRA, QLoRA, PEFT, TRL, Axolotl, vLLM, and a vector database (Pinecone or Weaviate). That skill stack is not rare in an absolute sense. What is rare is a recruiter who can identify the stack from a model card instead of a resume.

Senior generative-AI roles are running a 54-day time-to-fill against 38 days for mid-level AI/ML. Fourteen of those extra days are spent screening rebranded SWEs out of the funnel. Sourcing from HF profiles inverts that: the artifact is the screen.

Upload means they did the work. Download means they read about it. Most recruiter funnels filter out the entire pool worth hiring.

How to actually source from Hugging Face

Sourcing from Hugging Face is a two-step move: read the artifact, then reference it in the outreach. Both steps are where most recruiters give up, and both are the reason the channel still works.

Step 1: read the artifact

Open a model card and look for these six things:

• Base model. Llama-3, Qwen-2.5, Mistral, Gemma. Tells you what ecosystem they work in.
• Adapter type. LoRA rank and alpha, QLoRA quantization, full fine-tune. Tells you their compute constraints.
• Training framework. TRL, Axolotl, Unsloth, raw transformers. Tells you their toolchain preferences.
• Dataset. Public benchmark, custom scrape, synthetic. Tells you their taste in data.
• Eval. Reported numbers on a standard benchmark, or absence of eval. Tells you their rigor.
• License. Apache 2.0, MIT, non-commercial. Tells you whether they think about downstream use.

If the model card has all six, you are looking at somebody who has shipped fine-tuning work end to end. That is the profile the 135.8% demand spike is chasing.

Step 2: reference the work in outreach

80% of developers say they are open to hearing about roles, per daily.dev's December 2025 State of Trust report, but 43% ignore recruiter outreach entirely. The gap is closed almost exclusively by referencing the specific artifact.

A message that mentions "your Qwen-2.5-7B LoRA on the medical-QA split, particularly the choice to freeze the embedding layer" will get a reply. A message that says "I saw your impressive AI background" will not.

This is where Refolk collapses the two steps into one: you describe the person in plain English, including the artifact signal, and get a ranked shortlist across GitHub, LinkedIn, and the open web without stitching together three tools. The Hugging Face profile becomes a source you can query the same way you query a job title, which is what a hugging face recruiter workflow has been missing.

Where the pool actually sits

The skill-verified US fine-tuning pool is concentrated in three metros and roughly nine employers. Anything outside that footprint is either a poach target or a false positive.

From Refolk's index of the 3,177-person US pool:

• Metros: San Francisco Bay Area, New York City, Seattle. In that order, and it is not close.
• Frontier-lab employers: Meta AI, Apple ML, Amazon (mostly AGI and Bedrock orgs).
• Applied-AI startups: Twelve Labs, fileAI, Distyl AI, Sandgarden.
• Communities where they actually engage: Hugging Face forums and Discord, MLOps World, NeurIPS, local AI meetups.

The European pool of 28 breaks down as: Paris (4), London (3), the rest scattered. Named employers include Isomorphic Labs, Spotify, BCG X, and Max Planck Institute. If you are trying to hire in Europe, you are not sourcing, you are negotiating against four specific companies.

The founder version of this problem is even sharper. If you are hiring your first fine-tuning engineer for a seed-stage lab, the ~40 profiles that match your exact stack are probably split across three of the employers above, and the outreach has to reference the work by the second sentence. That is the exact gap Refolk closes for AI engineer sourcing at the seed and Series A stage: plain-English queries against the artifact-verified pool, not a title filter that returns Paychex.

The 12-to-18 month window

The recruiter-usage gap on GitHub (31%) and Hugging Face (lower) will not last. Every hiring guide published this quarter names HF as a channel, and the tooling to search it programmatically is arriving fast.

Three things happen when this becomes mainstream:

1. The fine-tuning premium compresses. The 25 to 40% band above median exists because the supply signal is invisible. Making it visible narrows the band.
2. Time-to-fill on senior gen-AI roles drops. The 54-day average is inflated by screening cost, not interview cost. Artifact-first sourcing cuts screening.
3. The 28-profile European pool gets picked over completely. US companies that want European fine-tuning hires need to move in the next two quarters or accept that the pool is gone.

Early-mover recruiters and founders benefit most in this window. That is the whole argument for building LLM fine-tuning hiring around Hugging Face sourcing now rather than in Q3 2027, when the daily.dev survey will read very differently and the 3,177-person US pool will have been contacted an average of nine times each.

FAQ

Is a Hugging Face profile really a better hiring signal than a LinkedIn title?

Yes, for LLM fine-tuning specifically. The LinkedIn "AI Engineer" title is 37% diluted at the pool level in the US (4,339 raw versus 3,177 skill-verified in Refolk's index), and the top employers in the raw cohort are non-AI companies. A Hugging Face model card with a specified base model, adapter type, dataset, and eval is a piece of proof-of-work that no title can approximate. For other AI roles like RAG engineering or ML infrastructure, LinkedIn plus GitHub is usually enough, but fine-tuning is the specialization where the artifact matters most.

How do I search Hugging Face without a dedicated tool?

You can filter models by base architecture, task, and library on huggingface.co directly, then click through to author profiles. It works for a first pass, but it does not join to LinkedIn employment history, does not deduplicate against GitHub identity, and does not rank by recency of upload. That last part matters because a 2023 fine-tune of an obsolete base model is a much weaker signal than a 2026 fine-tune of a current one. Purpose-built sourcing tools handle the join and the ranking; manual HF browsing does not.

What should the first outreach message actually say?

Reference the specific artifact by the second sentence, name a concrete technical choice they made, and ask a question rather than pitch a role. Something like: "Saw your Qwen-2.5-7B LoRA on medical-QA. Curious why you froze the embedding layer given the domain shift, most people don't. Working on something adjacent at [company], would you be up for a 15-minute call?" This clears the 43% ignore rate cited by daily.dev because it reads as peer-to-peer, not as a template. Recruiters who cannot personalize at this level should route sourced profiles to the hiring engineer for the first message.

Is 3,177 people really a big enough US pool to hire from?

For most companies, yes. The pool is roughly three times the size of the current annual demand for senior fine-tuning hires at frontier labs and applied-AI startups combined, and it is concentrated in three metros with mature engineering communities. The reason it feels small is that companies are competing for the same top 200 profiles, which is the same mistake the download distribution on Hugging Face itself makes visible. Widening the search into the long tail of contributors, engineers who have shipped one solid fine-tune but do not have a viral model, is where the actual hiring capacity lives.