Open-weight agent traces are runs of agents on models whose weights you run yourself, such as Qwen, DeepSeek, Kimi, GLM or Llama. Buyers favour them because the full reasoning is usually available and no hosted provider's output terms apply. Each model's own licence still does, and customer and employee data rules apply as for any trace. Check that your server and harness actually saved the reasoning.
- Open-weight runs avoid hosted providers' output clauses, but not the model's own licence.
- Read the licence of the exact model version that ran.
- Reasoning is returned in its own field on vLLM; old client code can read it as empty.
- Record model, version, serving stack, parser, harness and sampling settings for every run.
- Customer data, employee data and secrets rules apply exactly as for any trace.
Why do AI buyers prefer agent traces from open-weight models?
Two reasons. The full reasoning is usually there, and no hosted provider's clause on using outputs to build competing models applies.
Hosted APIs rarely return full reasoning. OpenAI's guide says "reasoning tokens are not visible via the API" (OpenAI), and Claude's docs say "the text in a thinking block is a summary of Claude's reasoning" (Claude docs). Full reasoning matters for training: Anthropic reports "A model trained on transcripts with summarized reasoning scored 71.8% on SWE-bench Verified, compared to 80.4% for a model trained on the full reasoning" (Anthropic). Runs on hosted models are not ruled out; counsel reads the provider terms first.
How common are open-weight models in production?
A growing minority. Menlo Ventures reports that "open-source models’ share of usage rose from 11% to 19% as well", while "Open-source remains 4% of spend despite 19% of usage" (Menlo Ventures).
The same report says "Chinese models grew to 89% of open-source spend, led by Kimi", followed by DeepSeek, GLM and Qwen. Open models are capable at agent work too: on an incident benchmark, "GLM-5.1 (Reasoning) leads open weights models at 40%" (Artificial Analysis).
Can I license agent runs made with DeepSeek, Qwen, Kimi or GLM?
Usually the model licence is not what stops you, but read the licence of the exact version you ran. The table quotes each licence as published on the model's page; it is a starting point, not a legal opinion.
| Family | What the licence says (exact words) | What to note |
|---|---|---|
| DeepSeek | R1: "licensed under the MIT License" and the series supports "distillation for training other LLMs" (DeepSeek-R1). V3.2: "This repository and the model weights are licensed under the MIT License" (DeepSeek-V3.2) | Check the card of the version you ran |
| Qwen | Qwen3-8B: "License: apache-2.0" (Qwen3-8B) | Some Qwen releases use other licences; check each size |
| GLM | GLM-5.1: "License: mit" (GLM-5.1) | Check the card of the version you ran |
| Kimi | Modified MIT: above a usage or revenue threshold, you "shall prominently display" the name Kimi K2 "on the user interface of such product or service" (Kimi K2 licence) | An attribution duty for very large products, set out in the licence |
| Llama | If you use "any outputs or results of the Llama Materials to create, train, fine tune, or otherwise improve an AI model, which is distributed or made available", the model's name must begin with "Llama"; a separate licence is needed above "700 million monthly active users" (Llama 3.1 licence) | A naming rule that reaches models trained on outputs |
| gpt-oss | "Permissive Apache 2.0 license" (gpt-oss-120b) | Open weights from a hosted-model provider; the API terms do not apply to self-hosted runs |
This page is general information, not legal advice. A fine-tuned model keeps its base model's licence terms; record the base model as well as your own version.
How do I capture the reasoning of an open-weight model served on vLLM or SGLang?
Turn on the server's reasoning parser, read the reasoning field your server version uses, and check that a sample run in your log actually holds reasoning text. An empty field usually means the parser was off, thinking was disabled or the client read the wrong field.
In vLLM, the reasoning-parser "flag specifies the reasoning parser to use for extracting reasoning content from the model output" (vLLM docs). The field was renamed from reasoning_content to reasoning, and the docs warn: "Otherwise, your client code could silently read an empty" field (vLLM docs). It happens in practice: one harness adapter read only the old field, so "every reasoning delta is silently dropped" and "no reasoning block reaches the session log or the UI" (deepseek-harness discussion). SGLang has its own parser settings; check its docs for your version.
If you log with OpenTelemetry GenAI, the output message schema has a part that "Represents reasoning/thinking content received from the model" (OpenTelemetry). Check that it is populated.
What should I record about the model and harness for each run?
Enough for someone else to know exactly what produced the run. Fill these in once per deployment and stamp each run with the deployment ID.
| Field | Example |
|---|---|
| model, model_version | Exact model name and revision from the model page |
| base_model, licence | For fine-tunes: the base model and its licence |
| serving_stack, version | vLLM or SGLang, with version |
| reasoning_parser, reasoning_field | The parser name and the field your client reads |
| harness, harness_version | LangGraph, Agno, OpenHands or your own, with version |
| sampling | Temperature, top_p, max tokens, thinking budget if set |
| reasoning_visibility | Full, summary or none, checked on a sample |
Example: a mixed fleet, labelled run by run
A fictional insurance claims team runs three agents. The figures are invented to show how to label a mixed set.
| Agent | Model | Runs in 6 months | Reasoning | Next step |
|---|---|---|---|---|
| Claims intake triage | Open-weight, self-hosted on vLLM | 4,200 | Full, after the client fix in month 2 | Describe months 3 to 6 first |
| Fraud referral review | Open-weight, self-hosted | 900 | Full | Strongest set: long runs with an investigator's decision |
| Customer letters | Hosted proprietary model | 11,000 | Summary only | Counsel reads the provider terms before anything else |
Fictional example. Labelling each run's model lets the open-weight runs move first while the hosted runs wait for counsel.
Common mistakes with open-weight runs
Most of them come from assuming open weights settle every question.
- Assuming an open licence clears customer data, employee data or secrets in the runs.
- Reading the licence of a different version or size than the one that ran.
- Not checking that reasoning reached the log.
- Losing track of which runs used which model in a mixed fleet.
- Leaving out the base model for a fine-tune.
- Switching models only to make traces more valuable. Choose models for the work and record what ran.
Open-weight agent runs as a data asset
Runs on open-weight models with full reasoning, a stated harness and a checked outcome are what buyers ask for first. They are also the quickest to clear, because the hosted-provider question does not arise.
Origin helps you sort runs by model, check the licence of each version, confirm reasoning was kept and describe the set. See also non-coding agent traces. We help you license your data. Origin Data Partners is an independent data partner: we tell you what your records could be worth and prepare the records, the terms and the license with you. The AI company that licenses them pays your company; a buyer may pay Origin when a deal closes.
Does your company hold records like these? See what your data is worth.
Common questions
Are open-weight agent traces always free to license?
No. The model licence, customer data, employee data and any secrets in the runs still apply. Open weights only remove the hosted provider's output terms.
Our reasoning field is empty. Why?
Usually the reasoning parser was not set, thinking was turned off, or the client read the old field name. On vLLM the field is now called reasoning; it used to be reasoning_content.
Do agent traces from Chinese open-weight models raise extra issues?
Check the model's licence and any customer contract or sector rule about where models come from. This is general information, not legal advice; your counsel decides.
Is a fine-tuned open model still open-weight for licensing traces?
Record the base model and its licence as well as your fine-tune. The base model's terms generally carry through to your version.
Do open-weight model licences allow training on their outputs?
The licences quoted here allow broad use, with conditions: Llama adds a naming rule for models trained on its outputs, and Kimi adds an attribution rule for very large products. Read the licence of the exact version you ran.
Sources and further reading
- OpenAI API docs: reasoning models
- Claude API docs: thinking
- Anthropic: distillation defenses (12 October 2026)
- Menlo Ventures: 2026 state of enterprise AI (3 November 2026)
- Artificial Analysis: ITBench-AA, SRE incident tasks (27 May 2026)
- DeepSeek-R1 model card (MIT License)
- DeepSeek-V3.2 model card (MIT)
- Qwen3-8B model card (Apache 2.0)
- GLM-5.1 model card (MIT)
- Kimi K2 Instruct: Modified MIT License
- Llama 3.1 Community License Agreement
- gpt-oss-120b model card (Apache 2.0)
- vLLM docs: reasoning outputs
- deepseek-harness discussion 199: reasoning deltas dropped against vLLM (13 August 2026)
- OpenTelemetry GenAI output message schema (reasoning part)
