Is post-training tooling worth building in 2026?

Is post-training tooling worth building in 2026?

Two unrelated GitHub repos spiked the same week in August and pointed at the same underlying question. Unsloth, the open-source library for LoRA and QLoRA fine-tuning, was climbing at roughly 580 new stars a day as of August 16, sitting at 72,455 total and API-verified by maybe worth building's radar that day. Five days later it's at 74,252, still climbing, just not at that spike's pace. Meanwhile mattpocock/skills, a folder of markdown instructions for coding agents that has nothing to do with training a model, went from 219,154 stars on August 16 (which we already covered here) to 229,280 by August 21, a gain of over 10,000 stars in five days with zero GPUs involved.

Those two numbers are the whole story. One tool changes a model's weights. The other changes nothing but a text file the agent reads before it acts. Both are being adopted at a scale most Series A startups would kill for, and only one of them costs a GPU-hour.

Why post-training tooling is and isn't a real opportunity

The case for it: post-training is real, it's growing, and it isn't going away. LoRA and QLoRA cut the cost of adapting a model from a full retrain to something that fits on a single consumer GPU, and that unlock created an entire category of use, from Character.AI-style behavior shaping to teaching a model to fix malformed hardware-description code, a real production use case one engineer described on Hacker News: "For tasks which involve generation of code or other non-natural language output, we've found that fine-tuning with the right dataset can lift performance rapidly and decisively." The demand is not manufactured.

The case against it: the category already ran its consolidation cycle, fast. Predibase, built by Google and Uber alumni around the open-source LoRAX project, raised a $12.2M Series A from Andreessen Horowitz and was acquired by Rubrik on June 25, 2025, in a deal reported between $100M and $500M. OpenPipe, which post-trains models on reinforcement learning from a builder's own workflows, raised a $6.7M seed and was acquired by CoreWeave on September 3, 2025, the same buyer that had already swallowed Weights & Biases months earlier. Two credible, funded, technically serious teams in this exact space, and neither one is still an independent company. That's not a market that's still forming. That's a market that already told you the standalone-product outcome doesn't clear, and the infrastructure layer absorbs you instead.

What does the market data say?

Run the numbers side by side. Unsloth: $500K raised total, one seed round, Y Combinator 2024, roughly 19 employees, and 74,252 GitHub stars carrying the entire category's mindshare on essentially no capital. Predibase: $12.2M raised, acquired for an estimated $100M to $500M after two years. OpenPipe: $6.7M raised, acquired for undisclosed terms after two years. The open-source project with the smallest check is winning the distribution war against the two that raised real venture money to compete on the same problem. That's the opposite of how most categories play out, and it's the single clearest signal that a standalone "make fine-tuning easier" company is fighting the wrong fight.

The skills side tells a related but distinct story. mattpocock/skills, google/skills, and obra/superpowers between them logged over half a million combined stars as of mid-August, per our own agent-skills-registry reporting. None of that is post-training. It's a cheaper, faster substitute for a chunk of what people used to reach for fine-tuning to solve: getting an AI system to reliably follow a specific workflow. When the free, no-GPU version of "teach the model to do X" explodes in adoption at the exact moment the weight-based version consolidates into two acquisitions, that's not a coincidence worth ignoring.

The companies already doing this: the receipts

• Unsloth. Open-source, MIT-adjacent, $500K raised, 74,252 stars. Claims 2x faster training with 70% less VRAM than standard fine-tuning, per its own repo. Monetizes through Unsloth Studio and enterprise support rather than gating the core library.
• Predibase (now Rubrik). Built LoRAX and Turbo LoRA, letting hundreds of fine-tuned adapters run on a single GPU. Raised $12.2M from a16z, acquired June 25, 2025, folded into Rubrik's agentic AI push.
• OpenPipe (now CoreWeave). Reinforcement-learning post-training via its open-source ART toolkit. Raised $6.7M seed, acquired September 3, 2025, the second post-training acquisition CoreWeave made inside a year.
• Axolotl. A second open-source fine-tuning framework with no funding round attached to it at all, and the tool a frustrated Unsloth user switched to mid-issue, described by that user as having "a very attentive support team." No press release. No round. Real adoption anyway.

Four real players, and the pattern across all four is the same: capital didn't buy defensibility here. The two that raised the most money exited into someone else's platform. The two that raised the least (or nothing) kept the users.

What would actually make post-training tooling worth building?

Not another training library. The wedge is visible in a single GitHub issue filed against Unsloth on March 21, 2026 by a user named mags0ft: fine-tuning a small model kept crashing with an out-of-memory error depending on the dataset's format, for reasons neither Unsloth's docs nor its maintainers could explain. Another user, mykeehu, piled onto the same thread describing a training run that "was supposed to be a one-and-a-half-hour session" turning into "a nearly four-hour training session" because of the same unexplained VRAM behavior. Twenty-one comments and almost two months later, mags0ft closed the loop: "Okay, know what? Nevermind. Nobody else seems to have this issue. I've documented enough. To anyone having the same problem: Switch to Axolotl, they have a very attentive support team."

Read that closely. The complaint was never that fine-tuning doesn't work. It's that when a run silently misbehaves, there's no tooling that tells you why, and the fix on offer is switching to a different free library rather than paying anyone to solve it. That's the actual open lane: run-level diagnostics that catch a VRAM spike before it becomes an OOM crash, regression checks between checkpoints so a new fine-tune doesn't quietly get worse at the thing the last one was good at, an eval-and-retrain loop that flags drift automatically, and cost prediction before a job burns the GPU-hours. Sell that to the teams doing post-training at real scale, the ones for whom a two-month unresolved GitHub thread is an outage, not a hobby project. Skip anything that competes head-on with a free library maintained by people who, per the numbers above, are already winning.

If you're deciding what to build your own product on top of, the same discipline applies one layer up: don't hard-wire to a single model or a single fine-tune, because the best model to build a startup on keeps depreciating under you. And if the shape of your idea is "own the trust and versioning layer the free tools skipped," that's the identical bet we found worth making in the agent skills registry verdict, just applied to weight updates instead of markdown files. The underlying tradeoff between owning weights and renting a frontier API, including when fine-tuning your own data actually pays for itself, is covered in full in whether it's worth building on open-source AI models.

The one way this verdict is wrong: if Unsloth or Axolotl ship their own hosted diagnostics and eval layer natively, the way registries keep threatening to absorb the trust layer around skills, the standalone wedge closes fast. Watch whether the open-source maintainers treat reliability tooling as core roadmap or as someone else's problem. Right now, per that March issue, it's the latter.

Frequently asked questions

Is post-training tooling worth building in 2026?

Not as a training library. Unsloth already has 74,252 GitHub stars on a total raise of $500K, and the two funded startups that tried to out-UX it, Predibase and OpenPipe, both got acquired by infrastructure companies within 14 months of each other. The open wedge is the reliability and diagnostics layer around post-training, where even Unsloth's own users hit unexplained failures with no clear fix.

What is post-training in AI, and how is it different from pre-training?

Pre-training is the original, months-long run that turns raw text into a base model, costing frontier labs hundreds of millions of dollars. Post-training is everything after: fine-tuning (full, LoRA, or QLoRA), reinforcement learning, and distillation that takes a base model and shapes it toward a specific behavior or task. Post-training tooling is the software layer that makes that second step manageable on a fraction of the compute.

Is Unsloth free, and how does it make money?

The core library is open-source and free. Unsloth is a San Francisco company founded in 2023 by brothers Michael and Daniel Han, backed by a single $500K seed round from Y Combinator in 2024 with about 19 employees as of mid-2026. It monetizes through Unsloth Studio, paid support, and enterprise deployment rather than gating the core fine-tuning code.

What happened to OpenPipe and Predibase?

Both got acquired rather than reaching an independent exit. Predibase, built around the open-source LoRAX project, raised a $12.2M Series A from Andreessen Horowitz and was acquired by Rubrik on June 25, 2025, in a deal reported between $100M and $500M. OpenPipe raised a $6.7M seed and was acquired by CoreWeave on September 3, 2025, the same company that had already bought Weights & Biases.

Is fine-tuning still worth doing in 2026, or is RAG enough?

Depends what you're trying to fix. RAG and long context handle knowledge injection better than fine-tuning does. Fine-tuning still wins for shaping behavior: enforcing a strict output format, teaching a narrow skill, or cutting inference cost by shrinking a task down to a small tuned model. Use the wrong tool for the job and you'll spend weeks of GPU time for a worse result than a better prompt would have gotten you for free.

Are agent skills the same thing as fine-tuning?

No. A skill, like the ones in mattpocock/skills, is a markdown file the agent reads at runtime. No weights change, no GPU is touched. Fine-tuning bakes a behavior into the model's parameters permanently. They compete for the same job, teaching a system to behave a specific way, but skills win when the behavior can be described in English and fine-tuning wins when it can't.

What's the biggest complaint people have about fine-tuning tools like Unsloth?

Not accuracy. Reliability and diagnostics. A GitHub issue filed against Unsloth on March 21, 2026 documented VRAM usage ballooning unpredictably depending on dataset format, ran for 21 comments and nearly two months, and ended with the reporter giving up and switching to a competing free library. The training math worked. The tooling that explains why a run is failing didn't.

What would actually make a post-training tooling startup worth building in 2026?

Sell the layer the open-source darlings don't bother with: run-level diagnostics that explain a VRAM spike before it OOMs, regression detection between checkpoints, automated eval-and-retrain loops, and cost prediction before a training job starts. Sell it to teams already committed to weight-based post-training at real scale, not hobbyists who'll route around a bug by switching to a different free library.