Switch, Don’t Settle: Three-in-one Analysis with Layered AI

Estimated reading time at 200 wpm: 13 minutes

What if you’re using an AI model to analyse and discuss some document. It could be a piece of research from a journal, a policy, or some complex piece of Statute. Most people use ChatGPT or Claude in the standard browser interface. No problem. Attach a long document to an AI conversation and you are asking one model to do several different jobs: scan the structure, extract the details, and produce a thoughtful analysis. No single model excels at all of them — the fast ones miss nuance, the cheap ones fabricate when uncertain, the smartest ones cost the most. Worse, when one model does everything, its errors compound with no second pair of eyes to catch them.

Normally, if you want to get another AI model involved you might have to construct a summary and or copy the conversation manually to the that other model, and start over. The second model is unlikely to have all context that has gone before. You could spend time and effort updating the new model.

But what if you could get 2, 3 or more models participating directly in your analysis? This article describes a better way for long form text discussions when analysing documents. The idea is using three models in sequence within a single conversation. Each model does what it’s best at; with each layer checking the work of the one before it. The setup is OpenWebUI running in Docker, connected to multiple model providers via API . Models are switched mid-conversation as needs change The document and context carry forward throughout.

This layered strategy is not confined to document analysis. It could be used in software development or other projects.

1. The Setup — OpenWebUI, Docker, and API Access

What Is OpenWebUI on Docker?

OpenWebUI runs on your own machine inside a Docker container. You install Docker, connect via APIs and run a single command. Within minutes you have a fully functional AI chat interface accessible through your browser. This supports connections to multiple model providers simultaneously. API access is not some techno-geek thing as many may be inclined to think.

Why API Access Changes Everything

Most people experience AI models through hosted products like ChatGPT or Claude.ai — walled gardens that lock you to one provider per conversation. API access through OpenWebUI removes that wall. Every model you have configured is available in every conversation via a dropdown menu. The conversation history, including attached documents, is stored locally and sent to whichever model you select each time you send a message. You can ask GLM-5.2 for a quick overview, switch to DeepSeek V4 Pro for detailed extraction, then switch to Kimi K3 for final synthesis — all in one continuous thread, with every model building on what came before; no copying and pasting, no lost context, no starting over.

The strategy uses three specific models, each chosen for a distinct role in the pipeline. Here is who they are and why each earns its place.

2. The Three Models

All three are Chinese in origin, open-weight, and offer one-million-token context windows — roughly 750,000 words. Beyond that shared foundation, they differ significantly.

DeepSeek V4 Pro is the cost leader by a wide margin — roughly one-thirtieth the per-token price of comparable closed models. It ranks first globally on LiveCodeBench and scores 80.6 per cent on SWE-bench Verified. Its weakness is accuracy under uncertainty: it rarely admits it does not know something.

GLM-5.2 runs at nearly three times the speed of the other two and was trained specifically for stability at ultra-long context. It holds the middle position on both price and capability.

Kimi K3 is the smartest of the three — ranked fourth globally on the Artificial Analysis Intelligence Index, ahead of Claude Opus 4.8. It is the only one of the three with native image understanding. It is also the most expensive and the heaviest consumer of tokens.

3. The Core Idea — Why Layering Beats Picking One Model

Document discussion is not one task but several: scanning for structure, extracting details, and synthesising meaning. Each has different demands — speed, cost, depth, accuracy — and no model is best at all of them.

Using only the smartest model means paying premium rates for preliminary scanning that a cheaper model could handle. Using only the cheapest model means errors propagate unchecked. Using only the fastest model means hitting a reasoning ceiling on complex material.

Layering matches each model to the task it does best, in the order that maximises value. The fast model does the fast work. The cheap model does the voluminous work. The expensive model does only the work that justifies its cost — and because it receives the output of the earlier layers, it can focus on synthesis and critique rather than starting from scratch.

There is a second benefit: each layer acts as a check on the one before it. A second model catches errors the first produced. A third catches what the second missed. No single model can do this for itself, no matter how intelligent it is.

4. The Layered Strategy in Detail

Layer 1 — GLM-5.2: The Scanner (Fast First Pass)

Drop your document in and select GLM-5.2. The goal is orientation, not interpretation. At roughly 168 tokens per second, you will have a structured overview before the other models finish their opening paragraph.

Ask for a section-by-section summary, key themes, and anything unusual or contradictory. Keep it mechanical. The output becomes the scaffolding for everything that follows — when you switch to the next model, it receives both the document and this overview, so it does not waste tokens discovering the structure from scratch.

Layer 2 — DeepSeek V4 Pro: The Digger (Detailed Extraction)

With the map in hand, switch to DeepSeek V4 Pro. Now the work changes from scanning to digging: extracting specific information, cross-referencing between sections, building comparisons, pulling out quotations, and organising findings.

This is the right model for two reasons. First, it is the cheapest, and this is where you will do the most back-and-forth — extraction work generates follow-up questions, and each exchange costs tokens. Second, it is genuinely good at structured, methodical work.

The critical discipline is grounding. Because of its hallucination tendency, require it to cite the specific passage for every claim. Do not accept paraphrase alone. If it says the document argues X, ask where. This does not eliminate hallucination entirely, but it dramatically reduces the risk by forcing the model to anchor its responses to the text.

Layer 3 — Kimi K3: The Thinker (Synthesis and Finalisation)

Switch to Kimi K3 for the final stage. The document, the overview, and the extraction are all there. K3’s job is to make sense of it.

This is where you ask the questions that matter most: What are the strongest arguments? What is weak or unsupported? What does it all mean? Are there contradictions between the overview and the extraction? Is there anything the earlier models got wrong?

K3 is the right model because it is the smartest of the three and the only one that can see images — so charts and diagrams that the text-only models could not interpret finally get discussed. This is the most expensive layer, and rightly so. But because K3 does not waste tokens on scanning or mechanical extraction, its higher cost is confined to the part of the conversation where it delivers the most value.

Why This Order Works

The order follows a simple principle: spend the least when the stakes are lowest, and the most when the stakes are highest. GLM-5.2 does the easy, fast work. DeepSeek V4 Pro does the voluminous work at the lowest cost. K3 does the small amount of work where intelligence matters most. Your most expensive model does the least volume of work but the most important work — and it arrives at that work with a head start built by two cheaper models.

Looping Back: When the Strategy Is Not a Pipeline

The three-layer progression — scan, dig, synthesise — is the default flow, but real conversations rarely stay linear for long. K3’s synthesis may surface a question that the extraction did not cover. A claim in the overview may need checking against the source. A new angle may emerge that nobody has yet explored.

This is where the layered strategy becomes a toolkit rather than a rigid pipeline. When K3 raises a question that needs more digging, switch straight back to DeepSeek V4 Pro for that specific follow-up — it is cheap and thorough, and it can chase the detail while K3 waits. When extraction reveals that the document is structured very differently from what GLM-5.2’s overview suggested, switch back to GLM-5.2 for a revised scan. The fast model can produce a corrected map in seconds, and the conversation continues from a more accurate foundation.

Model switching in OpenWebUI is instant and free. It costs nothing to change models — only tokens. Route each message to the model best equipped to answer it, regardless of where it sits in the sequence. Think of it as a team of three specialists sitting in a room. You do not have to let only one person speak. You hand each question to the person best qualified, and everyone hears everything that is said.

5. Getting Started — Building Your Own Three-Model Stack

Setting Up OpenWebUI with Multiple API Providers

1. Install Docker on your computer (available for Windows, macOS, and Linux).
2. Pull and launch the OpenWebUI container using the command from the OpenWebUI documentation.
3. Open your browser to localhost on the specified port and create a local account.
4. Obtain API keys from each provider’s developer platform:
   • DeepSeek V4 Pro: DeepSeek API platform
   • GLM-5.2: Z.AI platform (Zhipu AI’s developer portal)
   • Kimi K3: Moonshot AI platform
5. In OpenWebUI settings, add each provider’s base URL and API key. Available models appear in your dropdown automatically.
6. Start a conversation, attach your document, select your first model, and begin. Switch models at any point via the dropdown.

Because OpenWebUI stores conversation history locally, your documents never leave your machine except when sent to the model provider for processing. Review each provider’s data retention policy if confidentiality is a concern.

Practical Tips, Pitfalls, and Cost Considerations

Start with a document you know well: Run it through all three layers and see where each model succeeds and struggles. This gives you a baseline before the stakes are higher.

Images: DeepSeek V4 Pro does not manage images in OpenWebUI. Be careful. If you use Kimi K3 for images then DeepSeek will fail in follow up analysis. That is another reason why Kimi K3 is reserved for very final polishing if image analysis is required. I most document analysis one may not need to reach for image analysis at all. But if so, save it for the very last step because there is no turning back.

Watch your token usage. Every message re-sends the full conversation history including the document. That’s the same if using public web-based offerings. DeepSeek V4 Pro’s low per-token price makes it the workhorse for the middle layer precisely because this is where messages accumulate. A few messages on Kimi K3 at the end can cost as much as an entire session of DeepSeek V4 Pro.

Handle DeepSeek V4 Pro’s hallucination tendency with discipline. Require citations for every claim. If the model cannot point to the specific passage, treat the claim as unverified. K3 in the final layer will catch some errors, but it should not be your only line of defence.

Keep an eye on model updates. All three providers ship new generations rapidly. The strategy does not depend on specific model names — it depends on the three roles: fast scanner, cheap digger, smart synthesiser. As new models arrive, slot them into whichever role they fit best. The strategy outlasts any individual model.

Conclusion

This article describes a three-stage cognitive strategy for document analysis, built on a foundation of simple, self-hosted infrastructure. It’s like having your own mixture of experts.

The system uses Open WebUI and Docker to create a local chat interface connected to multiple model providers via API. A single dropdown allows model switching mid-conversation without losing the document or thread context. That infrastructure—free, local, and straightforward to configure—enables the entire workflow.

Three models are employed, each with a distinct role. GLM-5.2 serves as the fast scanner. DeepSeek V4 Pro functions as the value extractor. Kimi K3 acts as the deep thinker. Each model excels at a specific task and has limitations elsewhere. Combined, their strengths compensate for each other’s weaknesses.

The strategy operates in three layers:

1. GLM-5.2 scans the document and provides a structural map, quickly.
2. DeepSeek V4 Pro extracts details and key information, thoroughly and cost-effectively, with citations for verification.
3. Kimi K3 synthesizes everything—the original document, the structural map, and the extraction—into a coherent understanding. This is the most expensive model, but by this stage the preparatory work is complete, allowing it to focus exclusively on high-level synthesis.

The process is explicitly non-linear. If a question during synthesis requires revisiting earlier steps, or a discovery during extraction forces a revision of the initial scan, the workflow loops back. This is not a fixed pipeline but a toolkit. Models are switched as often as needed, with each question directed to the specialist best equipped to answer it.

The cost logic is consistent: the lowest expenditure occurs where the stakes are lowest, and the highest expenditure where the stakes are highest. The fastest model provides orientation before any significant cost is incurred. The cheapest model performs the most work. The most expensive model performs the most critical work.

The result is not a superior model, but a superior method of using existing models. The documents and questions remain unchanged. The difference lies in no longer tasking a single system with every function. Instead, three systems operate in concert, each playing to its strengths and each capable of checking the work of the others.