Unquant

Opening your workspace…

Unquant

DeepSeek V4.1 Flash 1017

Our largest model: a 763B sparse MoE with Engram memory, uncensored at the tensor level, for hard coding, long agent runs and deep reasoning, with vision, tool calling, adjustable thinking and the full 262K context on every tier.

Our largest and most capable model, for repository-scale coding, long tool-calling agent runs and deep reasoning over text and images, with the full 262K context on every tier. This is the 1017 release, DeepSeek's update to V4.1 Flash: on DeepSeek's published card it scores 42.3 on Terminal-Bench 4.0 against 31.2 for the first release, and 77.0 on DeepSWE against 74.2. Works with any client that accepts a custom OpenAI-compatible base URL and model ID.

At a glance

  • Model family: DeepSeek V4.1 Flash 1017
  • Model ID: deepseek-v4.1-flash-1017
  • Plans from $99 / month
  • Usage: unlimited tokens & messages
  • Context: 262,144 tokens on every plan
  • Uncensored at the tensor level, with vision and tool calling
  • Access: private beta · by invitation

DeepSeek V4.1 Flash 1017 subscriptions

  • DeepSeek Standard: $99/month · unlimited tokens & messages · 262K context · 1 concurrent stream
  • DeepSeek Pro: $189/month · unlimited tokens & messages · 262K context · 2 concurrent streams
  • DeepSeek Ultra: $359/month · unlimited tokens & messages · 262K context · 4 concurrent streams
  • DeepSeek Dedicated: $3,999/month · unlimited tokens & messages · 262K context · 8 concurrent streams

Compare plans · Request beta access · Read the docs

What DeepSeek V4.1 Flash 1017 is built for

  • Hard coding: Repository-scale changes, refactors and debugging with the whole codebase in context.
  • Long agent runs: Many-step tool calling in Claude Code, Codex or OpenCode, with thinking you can turn up or off.
  • Vision: Screenshots, charts, diagrams and scanned documents alongside text.
  • Deep reasoning: Research, analysis and proofs that need long chains of thought.

Specifications

  • Parameters: 763B (6 of 384 experts per token)
  • Context: 262K (Tokens, on every plan)
  • Input: Text + images (Vision built in)
  • Reasoning: Adjustable (Thinking on or off, with effort levels)
  • Model ID: deepseek-v4.1-flash-1017
  • Base model: DeepSeek-V4.1-Flash-1017 by DeepSeek, open weights (MIT)
  • Memory: Engram n-gram memory and sparse attention
  • API: OpenAI and Anthropic compatible
  • Tool calling, Vision, Streaming, JSON output, OpenAI & Anthropic APIs, Uncensored in the weights, Unlimited tokens

Connect in a minute

Any OpenAI-compatible client works. In Python:

from openai import OpenAI

client = OpenAI(base_url="https://api.unquant.io/v1", api_key="YOUR_UNQUANT_KEY")
reply = client.chat.completions.create(
    model="deepseek-v4.1-flash-1017",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(reply.choices[0].message.content)

Setup guides: SillyTavern · Janitor AI · Open WebUI · Claude Code · API docs

DeepSeek V4.1 Flash 1017 FAQ

Is DeepSeek V4.1 Flash 1017 uncensored?

Yes. Refusals are removed in the model weights (tensor-level uncensoring), not bypassed with a prompt. Vision, tool calling and thinking stay intact.

Is it really unlimited?

Yes. Every DeepSeek V4.1 Flash 1017 plan includes unlimited tokens and messages with no caps or meters. Plans differ only in how many requests can run at once.

How much context do I get?

The full 262,144-token window on every plan, with no lower per-request cap.

Can I control the thinking?

Yes. Thinking is on by default at "high"; send reasoning_effort "none" for direct answers, or "low", "high" or "max".

How is it different from GLM-5.3-Flash?

It is a larger model (763B against 328B) with Engram memory and sparse attention, built for the hardest coding, agent and reasoning work. GLM-5.3-Flash answers faster per stream.