Choosing a Model

Reveal SDK AI works with any supported provider and the models each one offers. Your choice of model is the single biggest factor in the quality, speed, and cost of the AI features you expose to users. This page gives concrete recommendations — a flagship cloud model and a self-hosted local model — backed by Reveal's ongoing benchmarking.

What the model has to do

Reveal asks the model to handle two very different kinds of work, and a model that excels at one is not guaranteed to excel at the other:

• Dashboard generation — the Chat endpoint turns a natural-language request ("show me sales by region for 2023") into a dashboard. This is the harder task: the model must emit strictly structured JSON that references real tables and fields, applies the correct chart types, and respects filters and rules. Small formatting mistakes cause the whole result to fail.
• Data insights — the Insights endpoint analyzes the data behind a visualization to produce summaries, statistical analysis (correlation, trend, mean/standard deviation), and forecasts. The output format is simpler, but the reasoning is more open-ended.

Recommended flagship models

In our latest runs the following models strike a great balance between quality and cost, being our top recommendations for general use cases from each major commercial provider:

• OpenAI GPT 5.6 Luna (OpenAI)
• Claude Haiku 4.5 (Anthropic)
• Gemini 3.5 Flash (Google Gemini)

ASP.NET

Program.cs

builder.Services.AddRevealAI()
  .AddOpenAI(options =>
  {
    options.ApiKey = builder.Configuration["RevealAI:OpenAI:ApiKey"];
    options.Model = "gpt-5.6-luna";
  });

Node.js

server.js

revealAI.withOptions({
  defaultProvider: 'openai',
  settings: {
    openai: {
      ApiKey: process.env.OPENAI_API_KEY,
      Model: 'gpt-5.6-luna'
    }
  }
});

Java

Application.java

Map<String, Object> aiSettings = Map.of(
  "openai", Map.of(
    "ApiKey", System.getenv("OPENAI_API_KEY"),
    "Model", "gpt-5.6-luna"));

See the corresponding provider help page for all configuration options.

Strong alternatives

If you are handling more complex tasks, you an also consider these other models that have obtained perfect scores in our benchmarks:

• OpenAI GPT 5.6 Sol (OpenAI)
• Claude Opus 4.7 (Anthropic)
• Claude Opus 4.8 (Anthropic)

Large reasoning models are not always better

For these structured tasks, dedicated reasoning models (like GPT 5.4 Pro and GPT 5.5 Pro) may be slower and less reliable than other, faster flagship models. They can take much longer to answer and are more prone to timeouts. Reach for a faster flagship model first.

Recommended local model

If you need to keep data on your own infrastructure or avoid per-token API costs, you can run a model locally and point the OpenAI provider at it through a custom endpoint.

Use Gemma 4 26B (gemma-4-26b-a4b-it). Among locally hosted models it delivers the best overall accuracy on both workloads, and it is small enough to run on a capable workstation (validated on a 64 GB Apple Silicon machine). Note: this model was recently updated after release to fix issues with the embedded prompt template, make sure to use the latest version for best results.

If your primary use case is Dashboard Generation, Qwen3.6 35B-A3B is a great choice, but its performance on Data Insights is significantly worse than Gemma 4.

If you want faster responses and a smaller memory footprint, GPT-OSS 20B is a strong runner-up — noticeably quicker, though a step behind Gemma 4 26B.

Run either model with a server that exposes an OpenAI-compatible API — such as Ollama, LM Studio, or vLLM — and set the Endpoint and Model options:

ASP.NET

Program.cs

builder.Services.AddRevealAI()
    .AddOpenAI(options =>
    {
        options.ApiKey = "ollama";                        // local servers ignore the key
        options.Endpoint = "http://localhost:11434/v1";
        options.Model = "gemma-4-26b-a4b-it";
    });

Node.js

server.js

revealAI.withOptions({
    defaultProvider: 'openai',
    settings: {
        openai: {
            ApiKey: 'ollama',                             // local servers ignore the key
            Endpoint: 'http://localhost:11434/v1',
            Model: 'gemma-4-26b-a4b-it'
        }
    }
});

Java

Application.java

Map<String, Object> aiSettings = Map.of(
        "openai", Map.of(
                "ApiKey", "ollama",                       // local servers ignore the key
                "Endpoint", "http://localhost:11434/v1",
                "Model", "gemma-4-26b-a4b-it"));

Local models and dashboard generation

Local models are a good fit for data insights and for straightforward dashboard generation. For complex, multi-widget dashboards, the strongest cloud models are still meaningfully more reliable — validate against your own prompts before committing to a fully local setup. Avoid "thinking"/reasoning local models: they struggle to stay within the required JSON schema and are slow on modest hardware.

Benchmark summary

The recommendations above come from Reveal's ongoing benchmarking. Each model runs two 10-task suites built on the Northwind sample database — one for dashboard generation, one for data insights — with tasks of increasing difficulty graded by an independent LLM judge. The score is how many of the 10 tasks the model handled reliably (a correct result on the majority of attempts), out of 10. DNF means the suite did not complete — the model failed the early tasks and the run was stopped. DNR means the run was not completed because the model was not practical to benchmark interactively in this workflow.

The full results for every model tested are below. Find your current model to see how it compares — if it scores poorly on the workload you care about, the models recommended above are drop-in replacements.

Cloud models

Model	Provider	Dashboard generation	Data insights
GPT-5.6 Luna (recommended)	OpenAI	10	10
GPT-5.6 Sol	OpenAI	10	10
GPT-5.6 Terra	OpenAI	9	10
GPT-5.5	OpenAI	10	10
GPT-5.5 Pro	OpenAI	DNR	DNR
GPT-5.4 Pro	OpenAI	DNR	DNR
GPT-5.4	OpenAI	6	10
GPT-5.4 mini	OpenAI	6	9
GPT-5.4 nano	OpenAI	6	6
GPT-5.3 Codex	OpenAI	7	9
GPT-5.2	OpenAI	8	9
GPT-5.1	OpenAI	8	8
GPT-5	OpenAI	6	6
GPT-4.1	OpenAI	10	9
Claude Fable 5	Anthropic	10	10
Claude Opus 4.8 (recommended)	Anthropic	10	10
Claude Opus 4.7 (recommended)	Anthropic	10	10
Claude Opus 4.6	Anthropic	9	10
Claude Opus 4.5	Anthropic	8	8
Claude Opus 4.1	Anthropic	10	4
Claude Sonnet 5	Anthropic	8	10
Claude Sonnet 4.6	Anthropic	6	10
Claude Sonnet 4.5	Anthropic	9	9
Claude Haiku 4.5 (recommended)	Anthropic	9	10
Gemini 3.5 Flash (recommended)	Google	10	10
Gemini 3.1 Pro	Google	8	9
Gemini 3 Pro	Google	7	9
Gemini 3 Flash	Google	6	7
Gemini 3.1 Flash Lite	Google	4	7
Gemini 2.5 Pro	Google	6	8
Gemini 2.5 Flash	Google	5	6
Mercury 2	Inception Labs	4	5

Newer and preview model versions were sometimes captured under early-access conditions and their scores may not reflect current generally-available behavior. Latency varies widely too: reasoning-oriented models (several of the GPT-5.x variants, and Claude Opus/Sonnet when thinking is engaged) answer much more slowly — especially on data insights — without a matching gain in reliability.

Open-weight models

These models can be self-hosted or accessed through a hosted-inference provider, using the custom endpoint configuration.

Model	Dashboard generation	Data insights	Notes
Gemma 4 26B (recommended)	8	9	Best open-weight accuracy overall
Qwen3.6 35B-A3B	10	5	Excellent dashboard generation; weaker insights (tested with reasoning disabled)
GPT-OSS 20B	6	6	Fastest; light memory footprint
GPT-OSS 120B	5	8	Larger; strong on data insights
Kimi K2 Instruct	1	9	Excellent insights, weak generation
Qwen3 32B	6	2	Decent generation, weak insights
Llama 4 Maverick 17B	3	6	Weak generation, decent insights
Llama 3.3 70B	1	6	Weak generation, decent insights
Gemma 3 12B	2	4	Smallest; limited on harder tasks
Qwen3.5 35B-A3B	3	DNF	Insights suite did not complete
Nemotron 3 Nano 30B	0	3	Reasoning model; poor schema adherence
Gemma 3 27B	DNF	—	Generation suite did not complete
Llama 3 70B	DNF	—	Generation suite did not complete

Open-weight scores depend heavily on quantization, host, and sampling settings — treat them as indicative of a well-configured run rather than fixed numbers. The results above use 4-bit to 8-bit quantizations on workstation-class hardware or a hosted-inference provider.

A moving target

Model quality shifts every time a provider ships a new version, so treat these figures as a snapshot rather than a permanent ranking. The best way to choose is to test the top candidates against your own data and prompts before you commit.