LLM Providers

Post-processing runs your matching Custom Prompts rule over the transcribed text with an LLM. The default provider is configured in Settings → Post-Processing; a rule can also pick its own provider and model (Step 4 in the rule) and then runs on those. Without a usable model for the rule — none downloaded, no key saved, or the stage switched Off while the rule has no provider of its own — rules and transforms are skipped, the dictation is pasted as transcribed and the pill says why.

Three options

• Local Model (Built-in) — download a model inside Speakly and run it offline, without installing Ollama.
• Ollama (Local) — badged 100% Offline: "Private local LLM processing on your device".
• Your API Key — badged BYOK: you supply a key for one of the cloud providers below.

Built-in local models

Open Settings → Post-Processing → Local Model (Built-in). Under Available Models, download a model and select its radio button. Qwen3 4B (Qwen3-4B-Instruct-2507) remains the recommended default; Speakly may recommend the lighter model for your hardware. Existing selections are kept. Downloads resume the same way as Whisper models: an interrupted one continues from what it already received, also after restarting the app, a download that stops receiving data for 60 seconds is retried, and each file is checked against the SHA-256 checksum Hugging Face publishes before it is used. Cancelling discards the download.

• Qwen3 4B — Q4_K_M: the recommended option for everyday use; approximately 2.50 GB to download.
• Qwen3 1.7B — Q4_K_M: the lighter option for modest machines; approximately 1.11 GB to download.
• GPT-OSS 20B — MXFP4: the existing option for machines with more memory; approximately 12.11 GB to download.
• Qwen3.5 4B — Q4_K_M: Experimental, approximately 2.74 GB to download.
• Gemma 4 E2B — QAT Q4_0: Experimental, approximately 3.35 GB to download.

BYOK providers

Pick one under Provider:

• Groq — Ultra-fast inference
• OpenAI — GPT models
• Anthropic — Claude models
• Google Gemini — Fast & cheap
• Mistral — European, GDPR
• OpenRouter — 100+ models
• Cerebras — Ultra-fast

Cloud model lists are fetched from each provider

For BYOK, paste a valid key and Speakly asks the provider which models your account can use, then fills the dropdown from that answer. That list never replaces the model you saved: a saved model the provider does not list stays selected, as a row of its own, until you pick another — if the provider has really withdrawn it, pick a new one from the list. (Anthropic lists older models under a dated id, so a saved claude-haiku-4-5 selects the claude-haiku-4-5-20251001 row.) Retired models Speakly knows about are replaced at startup: Groq's former Llama, Mixtral and Gemma models move to openai/gpt-oss-120b or openai/gpt-oss-20b, Mistral Large to mistral-medium-latest, and Gemini 2.0 Flash to gemini-3.6-flash. A Gemini model that Google has since closed to your account falls back to gemini-3.6-flash when it is used. The built-in local models use the catalog described above.

Until a valid key is present the field shows a message instead of a dropdown:

• Add a valid API key to load available models. — no key yet
• Loading models… — the fetch is in flight
• No models found — check that your API key is valid. — the provider rejected the key

Defaults on a fresh install

• Ollama Endpoint: http://127.0.0.1:11434
• Model Name (Ollama): llama3
• OpenAI model: gpt-5.6-luna
• Groq model: openai/gpt-oss-120b
• Google Gemini model: gemini-3.6-flash
• Anthropic model: claude-haiku-4-5
• Mistral model: mistral-small-latest
• Cerebras model: gpt-oss-120b

OpenRouter has no default: pick its model from the fetched list once your key is in — until one is selected, the dropdown reads Choose a model… and post-processing through OpenRouter cannot run. Choosing OpenRouter during onboarding starts you on openai/gpt-4o-mini.

Google Gemini takes a Google AI Studio key (aistudio.google.com/api-keys) — the same key the Google transcription engine uses; saving it in either place fills in the other. The dropdown lists Gemini text models only (no TTS, transcription, image or robotics models). Gemini 3 models think before they answer; Speakly asks for low reasoning effort so a rewrite stays quick, and a saved model that Google has since closed to your account falls back to gemini-3.6-flash.

Using Ollama

1. Install Ollama from ollama.com.
2. Run ollama serve, then ollama pull llama3.
3. In Settings → Post-Processing, pick Ollama (Local).
4. Leave Ollama Endpoint at http://127.0.0.1:11434 and set Model Name to the model you pulled.
5. Click Test Connection. It reports Connected or Connection failed.