Skip to main content

​
DeepSeek (Recommended)

Uses SSE streaming with OpenAI-compatible tool-calling format. DeepSeek is Wolffish’s recommended default provider for agentic tasks. The V4 family delivers frontier-class reasoning and tool-use reliability at a fraction of what competing frontier models charge on output-heavy workloads — while matching or exceeding their agentic performance on multi-step tool chains. It’s also MIT-licensed, so you can self-host for $0 in API fees if you have the infra. Since August 2026 DeepSeek bills peak/off-peak (peak 01:00–04:00 and 06:00–10:00 UTC; off-peak is half price), which is why prices below read as ranges. Connecting DeepSeek auto-selects DeepSeek-V4.1-Flash — served under the plain id deepseek-flash, fast, cheap, and built for the tool-heavy turns most agentic work is made of. Since September 11, 2026 it is the whole Flash line: it replaced deepseek-v4-flash, deepseek-v4-flash-vision-exp and the other V4 Flash aliases, which DeepSeek still routes to it. V4 Pro remains the heavier reasoning tier, one pick away in the composer’s model switch. Best for: Agentic multi-step workflows, tool calling, research chains, cost-efficient daily automations.
If you’re setting up Wolffish for the first time and want one provider that does it all — reliable tool use, strong reasoning, fast responses, minimal cost — start with DeepSeek and leave it on the deepseek-flash default. It sees images too, so computer use and browser control work without switching providers. You can always add Anthropic or OpenAI later for specific use cases.

​
Getting an API Key

  1. Go to platform.deepseek.com
  2. Sign up or log in
  3. Navigate to API Keys and create a new key
  4. Paste it into Wolffish → Settings → Models → DeepSeek

​
Models

The retired Flash ids still work. deepseek-v4-flash, deepseek-v4-flash-vision-exp and the other V4 Flash aliases are routed by DeepSeek to V4.1 Flash, and Wolffish treats them the same way — same 1M context, same 64K output ceiling, same vision gate. They no longer appear in the catalog because there is only one model behind all of them.deepseek-v4-pro is on borrowed time. From September 14, 2026 DeepSeek routes it to V4.1 Flash and bills at the Flash rate, while the row keeps its own prices until a V4.1 Pro ships.

​
Vision

DeepSeek was text-only across its whole lineup until August 21, 2026, when an experimental vision Flash arrived. Since September 11, 2026 vision is simply part of the default: deepseek-flash carries the vision badge, so the agent genuinely looks — screenshots, attached photos and image_view crops reach the model as pixels, verified live against the API. In practice that means computer use and browser control work on DeepSeek without switching to another provider first. deepseek-v4-pro remains text-only — an image part is rejected outright — so on it the agent works from captions and tools rather than sight, and Wolffish’s vision gate makes sure pixels are only ever sent where they’re accepted.
Before v1.0.287 the app did not recognise the version-less id: a chat on deepseek-flash was treated as an 8K-context, 16K-output, text-only model rather than the 1M-context vision model it is, and images were stripped before sending. Update if you are pinning that id by hand.

​
Reasoning modes

How this model reasons is set in the model card beside the chat input — hover the model switch to preview the card, click to pin it open. The top chip row, labelled Thinking, is the control: pick a chip and it applies from your next message. Two separate ideas combine here:

​
Thinking — whether the model reasons

  • Off — the model answers immediately. Fastest and cheapest; ideal for simple, direct tasks.
  • On (the chip reads Normal) — the model first works through the problem in a dedicated reasoning pass before replying. Slower and uses more tokens, but markedly more accurate on multi-step, logical, or ambiguous tasks.

​
Effort — how hard it thinks

Only effort-capable models expose this; it applies once thinking is on.
  • High — standard reasoning depth. The right default for most agentic work.
  • Max — the model reasons longer and deeper for the hardest problems. More tokens and latency in exchange for higher quality on complex work.

​
The chip row

The active chip is tinted; the rest sit quiet. Each model shows only the chips it genuinely supports — a model that always reasons has no Off, a model with a single mode shows that one chip inert, and a model that can’t reason at all replaces the row with “Reasoning is not supported by this model.” Wolffish remembers your choice per model. (Before v1.0.236 this was a colour-coded brain button on the composer; it moved into the model card, beside the Single/Workflow row, so every model knob sits in one panel.) On DeepSeek: All V4 models support Off / High / Max — and since v1.0.251 the three settings genuinely differ. The effort field used to be sent one level too deep in the request, where nothing complained and nothing listened, so High and Max measured less than a third of a percent apart — two settings that were never sent. With the field in its correct place and measured live: Off thinks not at all, High thinks hard, and Max thinks roughly twice as long as High on a problem that rewards it.

​
Model Selection & Retries

Wolffish communicates with LLMs via ten native cloud providers, an aggregator (OpenRouter), and a local option, all using pure fetch() — no SDKs. Each provider has its own streaming format and tool-calling convention, which wernicke.ts normalizes into a single interface. Select your Brain model explicitly from the model switch beside the chat input — the model you choose is the one that runs. There’s no cascade or fallback order; for parallel work, switch the chat into workflow mode, where the master can run agents on any connected model, DeepSeek included. When a cloud Brain hits a transient error, thalamus retries the same model on a backoff schedule (it also uses net.isOnline() for instant offline detection). It does not route you to a different provider on failure.