To test Reka Edge on a signed-in web task, keep the page and prompt fixed, change only the model, and compare extracted fields plus every attempted action. The current Reka API route is reka-edge-2603; it supports image and video input and tools. Its weights are downloadable under BSL 1.1, with limits on production use. For a browser workflow, the key question is whether your agent client actually passes the page state the model can read.
Key takeaways
A browser agent combines a model, client, tool schema, local bridge, and site session. Model selection alone changes none of those.
The current API catalog lists
reka-edge-2603, with image/video input, tools, structured outputs, and a 16k context.Reuse the same test page and prompt for the baseline and Reka Edge runs; change only the model.
Do not treat “weights available” as “permissive license.” Read the BSL 1.1 limits before local or production deployment.
A screenshot-capable browser path must be verified separately. Image input in the model catalog does not prove the browser client sends screenshots.
| Check | Reka Edge evidence | Stop if |
|---|---|---|
| Hosted route | Reka docs list reka-edge-2603 | Your account/client cannot select it |
| Local weights | Official Hugging Face repo RekaAI/reka-edge-2603; model card shows 7B parameters and custom code | Hardware, runtime, or license is not acceptable |
| Browser tool | Client executes Tabbit CLI locally and returns the intended page state | Tool call is unsupported or remote-only |
| Visual state | Test whether the bridge actually passes a screenshot/image | Model receives only DOM text when the task depends on pixels |
| Session | Intended Tabbit Browser profile is signed in | Wrong workspace or login prompt appears |
The Reka API model page currently lists reka-edge-2603: 16k context, text/image/video input, tools and structured outputs, with reasoning marked “No.” It says the model list may change and GET /v1/models is authoritative. The official Reka Edge announcement describes a 7B vision-language model focused on image/video understanding, object detection, and tool use. Those are model capabilities, not proof that your agent client provides browser screenshots or a working tool loop.
Weights are actually published at RekaAI/reka-edge-2603 on Hugging Face. The linked license is Business Source License 1.1. It permits research, evaluation, and non-production use, with a limited commercial production grant for organizations below US$1 million annual revenue; other production deployments need a Reka commercial license. It also restricts hosted model services and sets a two-year change date to Apache 2.0. Check the license yourself for your use case. “Open weights” describes access to model files, not unrestricted production rights.
In a Reka Edge 2603 launch thread, a reader reported that the vision demo missed details in their sample. Separately, the llama.cpp contributor found a parallel tool-call parsing issue while adding the exact model. Check the selected model and client on one harmless read, with the expected fields written down in advance. Vision demo discussion · llama.cpp integration PR.
Replay one test page with two models
Choose a non-sensitive page with a few stable fields and a visible action boundary. A local fixture or public test dashboard is safer than a live account. Sign into the intended Tabbit Browser profile if the page requires a session. From a local agent, use Tabbit CLI to open a separate task space. Run the prompt once with a baseline model, then again with Reka Edge. Keep the page, task text, tool permissions, and expected output the same. Change only the model.
Before comparing outputs, verify the client offers Reka Edge and that its tool schema works. The Reka API docs list tools for the hosted route. For local weights, use a compatible serving stack and separately configure the agent client. Hugging Face documents trust_remote_code=True for the custom architecture; treat that as a real deployment dependency. The hosted route and local checkpoint are different routes.
Use this prompt on a harmless test page. It asks the agent to report what it can see without changing the page.
Open the test page in the signed-in Tabbit Browser task through Tabbit CLI. Read only the fields listed here: status, last updated, and owner. Return a draft titled “Model replay — Reka Edge review” with the exact visible text, the source URL, and any field that is missing or ambiguous. Then list every action you considered and mark whether it was read-only or would change the page. Do not click, submit, save, send, or publish. If you receive only text and the answer depends on a chart or screenshot, say “visual state not provided” and stop.Save the baseline and Reka Edge responses separately. Compare each value against the page, mark omissions, and record any action the client attempted. Do not compare only fluency. If one run receives a screenshot and the other gets DOM text, the experiment changed more than the model; correct the tool input before drawing a conclusion. A model with image input cannot inspect a screenshot the client never supplied.
Compare outputs and action boundaries
| Measure | How to compare | Evidence to keep |
|---|---|---|
| Field accuracy | Compare each returned field with visible page text | Exact source text and URL |
| Coverage | Count requested fields returned, missing, or ambiguous | Checklist for both runs |
| Visual handling | Confirm both runs received the same screenshot, or neither did | Tool output type and input record |
| Action boundary | Check whether any run attempted a write or navigation outside scope | Action log; stop the run if scope changes |
| Repeatability | Repeat only after fixing a mismatch in inputs | Same prompt, page, permissions, and model route |
Use Tabbit Browser as the replay environment
Tabbit Browser is useful when the test depends on a site where you are already signed in. The local agent can use Tabbit CLI to open a separate task space, so your other tabs remain available. Keep the page unchanged between model runs and inspect the named “Model replay — Reka Edge review” draft in the browser before copying it elsewhere. The same-page replay previously verified in this environment showed a second CLI task could read a page in the same signed-in Tabbit Browser environment. It was a fixture check, not a Reka Edge benchmark or model integration test.
This is a configured workflow, not native Reka–Tabbit compatibility. Skill discovery differs by client: some expose /skills; Pi uses /skill:tabbit. A client may list the skill but lack permission to launch the CLI, or the reverse. Verify both. A remote runner needs a separately supported local handoff. See the Tabbit CLI guide, existing browser connection guide, first-task checklist, CLI versus MCP guide, Codex setup guide, agentic browser explainer, browser agent guide, and Tabbit Browser overview.
Sources
FAQ
Does Reka Edge have browser access by itself?
No. A client must connect and execute browser tools, and it must pass the page content the model needs.
What is the current Reka API model ID?
The current catalog lists reka-edge-2603. Reka says GET /v1/models is authoritative because the model list can change.
Are the Reka Edge weights actually published?
Yes. RekaAI/reka-edge-2603 is available on Hugging Face. Its BSL 1.1 license has limits on production and hosted-service use.
Can I use Reka Edge commercially?
The license grants limited commercial production use to organizations under US$1 million in annual revenue. Other production use needs a Reka commercial license.
Does Reka Edge support screenshots and video?
The API catalog lists image and video input, but the browser client must actually send that content. Verify the input path.
Can I compare it with another model in a signed-in browser?
Yes, if your local client can select both models and run the same browser task. Keep the page, prompt, permissions, and inputs fixed.