Crukx review, pricing and verdict
Stores related model checkpoints as exact deltas from their parent model, so fine-tunes take less space to keep and move. Free Explorer, CLI in private alpha.
- Best for
- AI Platform Engineer
- Not for
- Solo developer needing a free self-hosted option
Should you use Crukx?
Crukx suits ML teams that keep dozens of full-weight fine-tunes of the same base model and pay to store and copy each one. It saves only what changed and rebuilds every file exactly. Its CLI is still a waitlisted alpha, and it saves storage, not GPU time.
Unified observability and optimization in one platform, removing the need to stitch tracing and cost-analysis tools together.
New entrant in a crowded space dominated by Langfuse, Helicone, and Arize. Enterprise positioning may carve a niche but no clear differentiated moat is publicly visible yet.
Also for
- ML Engineering Lead
- ML teams storing many full-weight fine-tunes of one base model
Show 2 more
- Platform teams paying to move checkpoints between clouds and clusters
- Researchers who need to know what a public model was built from
Also not for
- Early-stage startup that cannot engage a sales process
Main trade-off
No public documentation, pricing, or third-party validation makes it almost impossible to evaluate without a direct sales call.
Crukx pricing
| Plan | Monthly price | What it includes |
|---|---|---|
| Explore | Free | Model Explorer examples, Every benchmark with its method, Reproduce commands |
| Build (private alpha) | Free | crukx scan, push and pull, Runs locally, CPU only, Direct line to the core team |
| Pilot | Custom | Runs on your models and infrastructure, Written report on bytes, time and rebuild checks, Go / no-go criteria agreed up front |
Crukx publishes no paid price as of 9 November 2026. The Explore level (Model Explorer and benchmarks) is free with no account, and the Build level, the local CLI, is free while it stays in private alpha. A pilot on your own checkpoints is an engineering engagement scoped with the team, with go / no-go criteria agreed before it starts; no figure is given for it.
What each plan includes
| Feature | Pilot | Explore | Build (private alpha) |
|---|---|---|---|
| Account needed | Engagement | — | Waitlist |
| Model Explorer | ✓ | ✓ | ✓ |
| Push and pull your models | ✓ | — | ✓ |
| Written measurement report | ✓ | — | — |
Before you commit
- Enterprise licensing likely requires annual commitment with no monthly self-serve option
- Implementation and onboarding support for enterprise-scale deployments may carry additional fees
What Crukx does
Crukx is a model storage tool that keeps a fine-tuned checkpoint as a lossless delta from its parent model and rebuilds it byte for byte. On its measured Qwen fine-tune pair, a 29.5 GB model shrank to 8.2 GB. The Explorer is free and the CLI is in private alpha.
CRX lossless deltas
Stores a fine-tune as only what changed from its parent; on Crukx's measured fine-tune pairs that came to 2.4 to 4.3 times smaller.
Exact rebuilds
Every pull rebuilds the original files byte for byte and checks them with SHA-256, so model quality cannot change.
Local CPU CLI
Three commands, scan, push and pull, run on your machine without a GPU; the measured Qwen pair packed in 2 minutes 14 seconds on 32 CPU threads.
Model DNA (experimental)
Compares weights to name a model's likely parent; its index covers 103 models across families such as Qwen, Llama, Gemma and Mistral.
Public benchmarks
Seven model pairs in BF16 and FP8 measured against Hugging Face Xet and ZipNN, including the case where CRX does not help.
Missing integrations
- OpenAI SDK
- Anthropic SDK
- LangChain
- Slack
Most users expect these and they are not currently supported.
Capabilities
- Model storage
- Checkpoints
- Fine tuning
- Model lineage
- Mlops
- Compression
Strengths and limits
Strengths
- The claims are unusually checkable: every figure on the site links to its method and reproduce commands, and the losing case is published too.
- Rebuilds are verified exact, so adopting it carries no risk to model quality, only to storage workflow.
- Trying it costs nothing today, from the no-account Explorer to the alpha CLI.
Limits
- The CLI is a waitlisted private alpha, so production use means a pilot rather than a self-serve plan.
- A delta needs its parent model stored too; unrelated models get no saving.
- No paid price, SLA or compliance certification is published.
- The older observai.dev site still describes a previous product, which can confuse anyone who finds it first.
Crukx alternatives
| Tool | Price from | Free plan | Runs on | Rating |
|---|---|---|---|---|
| Crukx This page | -- | Yes | Web | 3.0 |
 Hugging FaceHugging Face hosts and serves models for everyone and dedupes storage by chunk. Crukx only stores and rebuilds related checkpoints, and its benchmarks say chunk dedup found almost no overlap between full fine-tunes. | $9/mo | Yes | Web | 4.0 |
 OvermindOvermind turns agent traces into datasets and fine-tunes. Crukx starts after that, keeping the resulting checkpoints small to store. | $120/mo | Yes | Web | -- |
 CodagCodag lowers what a coding agent pays to read tool output. Crukx lowers what it costs to keep many versions of a model. | $499/mo | Yes | Mac · Windows | 3.5 |
Also compared
Together AIChoose Together AI when you want managed fine-tuning and inference instead of running storage yourself.
Crukx facts
- Deployment
- Cloud
- Works offline
- No
- Minimum skill level
- Intermediate
- Price band
- Enterprise pricing
- Languages
- English
- Training
- Benchmarks with method
- Training-data policy
- For pilots, Crukx says your weights stay on your own infrastructure; the CLI runs locally.
Questions about Crukx
How much does Crukx cost?
Nothing yet for most users. The Explorer and benchmarks are free without an account, and the CLI costs nothing while the alpha runs. Teams that want CRX measured on private storage can ask for a pilot, but the company has not published what that costs.
Can I try Crukx without paying?
Yes. You can open the Model Explorer with no sign-in to see what a public model is built from, and you can join the waitlist for the alpha CLI, which is free while the alpha lasts.
Which tools compete with Crukx in 2026?
Its own benchmarks set CRX against Hugging Face Xet and ZipNN, which dedupe or compress model files. Fine-tuning platforms like Overmind or Together AI produce the checkpoints Crukx stores, rather than competing with it on storage.
Does Crukx make my models faster?
No. Crukx says CRX reduces the bytes you store and move, and that it does not claim training or inference speedups until it has measured them. The rebuilt model is identical to the original, so its speed and quality are unchanged.
Is Crukx still an LLM observability tool?
No. In August 2026 it described itself as enterprise LLM observability at observai.dev. By November its site at crukx.dev sells checkpoint storage and model lineage instead, and that older domain shows a model-optimization page with a waitlist.
About Crukx
Crukx Inc. has changed products more than once in 2026. Hokai first listed it in August as an enterprise LLM observability layer at observai.dev, and its own site now lives at crukx.dev with a different job: making related model checkpoints smaller to store and move. Its format, CRX, keeps a fine-tune as a lossless delta from the parent model it was trained from, then rebuilds the original files byte for byte, checked with SHA-256. On its measured Qwen2.5-14B to Instruct pair in BF16, a 29.5 GB model became an 8.2 GB delta, 3.61 times smaller.
Read the full background
There are three ways in. The Model Explorer and the benchmark pages are open with no account. The local CLI (crukx scan, push and pull) runs on a CPU and is in a free private alpha by waitlist. Teams with private checkpoints can request a pilot, where Crukx engineers measure CRX on their own storage and hand over a written report with the cases where it did not help. An experimental Model DNA feature compares weights to place a model in its family; Crukx reports it called families correctly 96.5% of the time across 515 public models.
Hokai files it under machine learning tools, where the fine-tuning platform Overmind produces exactly the checkpoints Crukx would store, and its benchmarks compare CRX with the Xet storage behind Hugging Face. The parent models it measures come from families like Qwen, built by Alibaba Cloud. To see how it differs from Codag, a coding-agent cost tool it used to be compared with, read Codag vs Crukx.
Who should skip it: anyone who keeps one or two models, or who stores LoRA adapters rather than full-weight fine-tunes, since the saving comes from many full copies of the same base. It also does not make training or inference faster; Crukx says plainly it reduces bytes stored and moved, not GPU time. A team that wants hosted fine-tuning and serving in one place is better served by Together AI or Fireworks AI.
Sources and guides
Official sources
- Crukx official site Vendor
- Crukx benchmarks and method Docs
- Work with Crukx Docs
- Crukx on GitHub Docs
- GitHub repository GitHub
Product facts on this page (pricing, plans, integrations, compliance) are read from Crukx's own pages and documentation. The verdict is Hokai's reading of them, not the vendor's claim.
Hokai guides covering Crukx
Also mentioned in
Lists that include Crukx
More AI Tools on Hokai
JevModelHosted console and API for the Jev decision model: typed Choice, Score and Noul answers, with a free token allowance to start.
GodelaGodela trains AI physics models on your simulation data and answers what-if queries in seconds. YC-backed, pilot access only.
fast.aifast.ai is a free nonprofit deep learning course reaching 500,000+ students with practical top-down teaching and open-source fastai library.AtmoAtmo builds AI weather forecasts up to 50% more accurate than traditional models, used by the US Air Force, Navy, and the Philippine government.