Getting the Most Speed Out of Your Device with AutoForge

What is AutoForge?

AutoForge is TokForge's built-in speed test and tuning screen. Think of it as a performance coach for your phone. It tests the settings that matter for the model you pick (processor cores, the CPU or the GPU paths your phone and model support, memory settings) and stages the best result it found. You review it and tap Apply. No guesswork.

AutoForge has three tabs: Tune runs the optimization, Results keeps your speed tests, and Saved setups holds the configurations you can apply again later.

Why Auto-Tuning Matters

Here's the reality: different chipsets perform very differently. The same model can run best on the CPU on one phone and on the GPU on another, and the best number of processor cores changes from chip to chip. Without measuring, you're just guessing.

AutoForge tests your specific hardware and finds the sweet spot. The impact is real:

• Picking the right compute path matters, and the right path is not always the GPU. On several device profiles in our own fleet the CPU path is the faster one, which is why AutoForge measures instead of assuming
• Thread tuning can add further measurable gains
• Memory settings such as conversation memory size and the KV cache mode change both speed and how much RAM the model needs

How to Run AutoForge (Step by Step)

1. Open AutoForge

1. Open TokForge on your Android phone
2. Tap the AutoForge tile on the Home screen, or open Settings → Speed & engine and tap the AutoForge card
3. You'll see three tabs: Tune, Results, and Saved setups

2. Check the Tune Tab

The Tune tab shows your phone's current setup:

• Model to tune: which model AutoForge will test
• Runs on: where that model runs now (CPU or GPU)
• What your phone can do: the speed-up features TokForge detected on this phone (see "Device Capabilities Explained" below)

3. Optimize My Phone

Pick how thorough the test should be (Instant, Quick, Thorough or Most thorough), then tap "Optimize my phone". TokForge will:

• Run a series of speed tests with different settings
• Test the settings that apply to your model:
  • Processor cores (thread counts)
  • Where the model runs: the CPU, or a GPU path your phone and model support
  • KV cache and attention settings, and conversation memory size
  • For MNN models: speed vs. quality (precision) and, on the Thorough and Most thorough tests where it is available, speculative decoding
• Duration: Instant stages known-good settings without running speed tests, Quick takes a minute or two, and Thorough and Most thorough take longer and use more battery
• Quick, Thorough and Most thorough run through TokForge's local developer API, so the app tells you to turn on Developer API server (Advanced) in Settings → Connections → Developer tools (Advanced) first. Instant works without it.

Pro tip: Run AutoForge after a fresh reboot for the cleanest results. Close all background apps and let your device settle for a minute before starting.

4. Review and Apply

When the test finishes, AutoForge stages the best result and shows a Before and After comparison. Nothing changes until you tap Apply; Rollback restores your previous setup at any time. Switch to the Results tab to see your speed tests.

Understanding the Results

Speed test results use a few key metrics. Here's what they mean:

The Results tab keeps your speed tests. Each result shows the model, the chip and where it ran. Tap Run Benchmark to test the loaded model, share a result as a card, or submit it to the public phone-speed leaderboard.

Device Capabilities Explained

The What your phone can do card on the Tune tab lists the speed-up features TokForge detected. Whether each one is used depends on the model you load and how it is set to run. Understanding these helps you interpret your results:

Fast CPU Math (KleidiAI)

ARM's KleidiAI CPU routines use the i8mm instructions found on Armv9 CPU cores; older cores such as the Cortex-A78 do not have them. On phones that support it, prompt processing (prefill) on the CPU gets faster, and GGUF models in Q4_0 format are marked as KleidiAI accelerated in the catalogue.

GPU Accelerated (OpenCL & Vulkan)

On some phones, some models can take a GPU path. Which GPU path is available depends on the chip and the model: OpenCL mostly on Snapdragon phones, Vulkan mostly for GGUF models on phones with other GPUs, such as the Pixel 9 and the Xiaomi 14T Pro. AutoForge compares the paths your phone and model support against the CPU, and stages the faster one. See the backends guide for details.

CPU Optimized

Works on every phone. No GPU needed. On many phones the CPU is the fastest path for chat, which is why AutoForge measures instead of assuming.

Saved Setups

The Saved setups tab holds the configurations your tunes produced. Run a tune from the Tune tab to create one. From there you can:

• Apply: switch back to a saved setup with one tap. If it runs the model on a different backend, TokForge asks you to reload the model.
• Export profiles: save your setups to a file.
• Import profiles: load setups from a file.

On the Tune tab, AutoForge can also show the best known config for the selected model, with its own Apply button.

Pro Tips for Best Results

• Reboot before auto-tuning. A fresh restart clears cached processes and gives you the clearest benchmark signal.
• Close background apps. Music, social media, system updates: they steal CPU and RAM. Close them before running benchmarks.
• Watch for thermal throttling. AutoForge records the phone's thermal state during a run and marks results taken while the phone was throttled. If that happens, let your phone cool down and run it again. Throttling skews results and real-world performance.
• Re-run after OS updates. Android updates sometimes change GPU drivers or kernel scheduling. Re-run AutoForge after major updates to catch new optimizations.
• Export detailed results. With the Developer API server turned on, GET /benchmark/export (with your access token) gives you your benchmark data in JSON for deeper analysis.
• Test with real workloads. Benchmarks show peak performance; real chat and reasoning workloads vary. Use saved setups as a starting point and fine-tune based on feel.

Next Steps

After running AutoForge, you have a config tuned to your device. Now explore:

• Model selection: How to choose between models for speed vs. quality (see our Backend Comparison Guide).
• Memory saving: How the optional TurboQuant setting cuts memory use for MNN models (see our TurboQuant Guide).
• API integration: How to run AutoForge and speed tests through the local developer API to build custom benchmarking workflows (see our API walkthrough).