> For the complete documentation index, see [llms.txt](https://assay.gitbook.io/assay-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://assay.gitbook.io/assay-docs/resources/hosting-report.md).

# Hosting report

Report v0 on who serves open-weight models, what they declare, and why the checks don't travel with the output.

**Where:** `report/REPORT_v0.md` and `report/index.html` in the repo. The raw data is `report/data/openrouter_endpoints_2026-09-26.json`.

Report v0 reads what hosts tell OpenRouter about the open-weight models they serve. Nothing in sections 2 and 3 of the report was tested. It shows what a buyer can and can't see from a listing.

## How it was measured

| Field    | Value                                                                       |
| -------- | --------------------------------------------------------------------------- |
| Source   | `GET https://openrouter.ai/api/v1/models/{model}/endpoints`, public, no key |
| Captured | 26 Sep 2026, 17:25 UTC                                                      |
| Scope    | 11 author namespaces, 140 open-weight models, 56 providers                  |
| Listings | 746, of which 735 distinct endpoints                                        |

## Declared precision

| Declared precision     | Endpoints | Share |
| ---------------------- | --------- | ----- |
| fp8                    | 285       | 38.8% |
| not stated (`unknown`) | 289       | 39.3% |
| fp4                    | 81        | 11.0% |
| bf16                   | 51        | 6.9%  |
| int4                   | 17        | 2.3%  |
| nvfp4                  | 5         | 0.7%  |
| mxfp4                  | 4         | 0.5%  |
| fp16                   | 3         | 0.4%  |

## Same model, many versions

| Model                           | Endpoints | Precision not stated | Declares 4-bit | Output $/M      |
| ------------------------------- | --------- | -------------------- | -------------- | --------------- |
| z-ai/glm-5.3                    | 37        | 16                   | 8              | $1.19 to $8.80  |
| z-ai/glm-5.2                    | 30        | 9                    | 8              | $1.54 to $8.00  |
| deepseek/deepseek-v4-flash-0731 | 30        | 11                   | 5              | $0.13 to $1.32  |
| openai/gpt-oss-120b             | 23        | 9                    | 5              | $0.17 to $0.95  |
| deepseek/deepseek-v4-pro-0813   | 22        | 11                   | 2              | $0.75 to $4.95  |
| moonshotai/kimi-k3              | 19        | 9                    | 7              | $9.04 to $22.50 |
| moonshotai/kimi-k2.6            | 19        | 5                    | 10             | $1.72 to $4.60  |
| qwen/qwen3.8-27b                | 16        | 7                    | 1              | $1.80 to $4.40  |

Only 311 of the 746 listings advertise logprobs, so most outside checks have to work from the output text alone.

## What others measured

| Who                         | Finding                                                                                             |
| --------------------------- | --------------------------------------------------------------------------------------------------- |
| Olly                        | DeepSeek V4 Flash at 90% GPQA on DeepSeek's API and 75% on DigitalOcean                             |
| Cline                       | GLM-5.2 at 74.2% on CoreWeave and 61.8% on OpenRouter's default routing                             |
| Moonshot K2 Vendor Verifier | Tool-call schema accuracy from 100% down to 86.91% across vendors                                   |
| Glama                       | Delisted every third-party host because "some of them are obviously lying about their quantization" |

The report lists each source with its date so you can check it.

## How to read it

1. A missing precision field is not evidence of a problem. It only means a buyer can't tell from the listing.
2. A 4-bit label does not automatically mean a worse model. One comparison found a provider declaring fp4 beating one declaring fp8.
3. That is why the next step is measurement. Report v1 will grade every GLM-5.3 endpoint against Z.ai's own endpoint with `harness/assay_probe.py`, with 95% intervals and raw logs.

## Where else this shows up

| Place                                                                    | What it covers                 |
| ------------------------------------------------------------------------ | ------------------------------ |
| [Welcome to Assay](/assay-docs/welcome/readme.md)                        | The headline numbers           |
| [Verifiers and grades](/assay-docs/how-it-works/verifiers-and-grades.md) | How measured grades go onchain |

{% hint style="warning" %}
The report names no provider as serving a degraded model. Any endpoint that scores below the reference in report v1 gets its logs and 7 days to respond before publication.
{% endhint %}

Next: [FAQ](/assay-docs/resources/faq.md)


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://assay.gitbook.io/assay-docs/resources/hosting-report.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.