DeepSeek V4.1 Flash 1017 (Max)

TL;DR Highlight

DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.

Who Should Read

A backend developer running LLM APIs who is weighing a cheap open-weights model for agent and coding workloads.

Core Mechanics

Intelligence Index 45, over twice the peer median of 19
- It scores 45 on Artificial Analysis Intelligence Index v4.3.2.
- The median for open-weights models of similar size is 19.
210.6 tokens/s with 1.09s TTFT
- This is measured on DeepSeek's own API, and the peer median is 77.6 tok/s.
- The peer median for time to first token (TTFT) is 1.71s.
$0.30 input and $1.20 output per 1M tokens
- Peer medians are $0.43 input and $1.68 output.
- Cache discount is 98%, and the average cost per Intelligence Index task is $0.26.
It is verbose
- It generated 210M output tokens while running the Intelligence Index.
- The peer median is 140M.
552B total, only 16B active (MoE)
- It has a 1M token context window and accepts text and image input with text output.

Community Reactions

Max effort is unrealistic, one commenter argues
- They note GPT 6.1 Sol (medium) is cheaper ($0.21 vs $0.26) and faster (2.9 vs 4.9 min) per task.
- They want results for other DeepSeek reasoning efforts to be added.
A month as a daily driver, per one user
- They ran tens of billions of tokens and found it capable on mid-level tasks versus proprietary subscriptions.
- On their internal evals 1017 is more token efficient but scores no higher than the previous version.
A 3-hour dev session cost about $0.40, per one user
- They sometimes ran 300k-token context requests because they did not compact often.
- It finished the work but needed more guidance than Claude would.
Weaknesses were reported
- One commenter says it still fails at reading sheet music.
- Another feels it trails GLM-5.3 on cyber, reverse engineering and hardware work.
Comparison with MiMo 2.6 Pro

How to Apply

Build a cost comparison table against your current model
- Pull your average input and output tokens per call from logs today.
- Plug in $0.30 input, $1.20 output and the 98% cache discount to compute monthly cost.
Cap output length for a verbose model
- Set max_tokens and add a conciseness instruction.
- Compare token count and accuracy against your old setup on 20 identical tasks.
Keep planning and review on a proprietary model and run execution on DeepSeek
- Split your agent pipeline into planning/review and repetitive execution steps.
- Swap only the execution step and measure the quality gap.
Benchmark reasoning effort levels on your own tasks
- The published numbers are max effort, so test both medium and max.
- Pick the effort whose quality justifies its time and cost.

Try It

copy & run

# Paste into Claude or ChatGPT: model switch cost calculation prompt

Here is our LLM usage:
- Current model: {{current model}}
- Calls per month: {{calls per month}}
- Avg input tokens per call: {{input tokens}}
- Avg output tokens per call: {{output tokens}}
- Cache hit ratio: {{0 to 1}}

The candidate is DeepSeek V4.1 Flash 1017 (Max). Pricing is $0.30 input and $1.20 output per 1M tokens, with a 98% cache discount. It is reportedly verbose compared to peers (210M vs a 140M median).

Compute monthly cost in a table. Show two cases: output tokens at 1.0x and at 1.5x.

Next step: paste your current model's price list and ask for the same table to compare.

Terminology

Intelligence IndexA composite score Artificial Analysis builds from several benchmarks. Think of it as the model's overall exam score.

MoE (Mixture of Experts)An architecture with many expert sub-networks where only a few work on each token. Here only 16B of 552B are active.

TTFTTime from sending a request to receiving the first token. Like the wait before the first dish arrives at a restaurant.

VerbosityHow many output tokens a model produces on the same tasks. More tokens mean higher cost and latency.

Cache discountA price cut on input when a previously processed prompt is reused. For this model it is 98%.