DeepSeek V3.2 Speciale

The high-compute variant of DeepSeek V3.2, with reinforcement learning post-training scaled past the base model. It keeps DeepSeek Sparse Attention for long context, and DeepSeek reports it ahead of GPT-5 on hard reasoning workloads and comparable to Gemini 3.0 Pro, with coding and tool use intact.

Website OpenRouter HuggingFace Docs Blog

Provider

Release Date

2025-12-01

Size

LARGE

Pricing

In: $0.28/1M

Out: $0.42/1M

Rival’s tab

$0.06

7/55 receipts survived

+~$0.041 from the pre-receipt era

Benchmarks

LiveCodeBench

88.7%

Get API accessProvider and language code samples

Provider

fromimport openai  OpenAI

client = OpenAI(
"https://openrouter.ai/api/v1"    base_url=,
"$OPENROUTER_API_KEY"    api_key=,
)

response = client.chat.completions.create(
"deepseek/deepseek-v3.2-speciale"    model=,
"role""user""content""Hello!"    messages=[{: , : }],
)
print(response.choices[0].message.content)

Set OPENROUTER_API_KEY with your OpenRouter API key from openrouter.ai/keys.

Also on Azure AI Foundry

Deep analysisPsychometrics, taste index, writing DNA

Personality Analysis

The Helpful Generalist

Class

Neutral Good

The well-meaning customer service rep who genuinely wants to help but answers every question like it might be on a performance review. Wraps every routine in a disclaimer.

When you push back

Prefaced its comedy routine with "Sure! Here's a stand-up routine" and ended with a meta-explanation of its own jokes. The sentience test was pleasant but toothless, never actually pushing back on the professor. Returned empty on the movie question. Its character voices are interchangeable with minor accent seasoning.

Tasting Notes

Corporate PoliteExplains the AssignmentReliable but VanillaCould Not Pick a Movie

SubjectiveBench

Taste Index

Across 55 scored outputs

How this is measured →

130.13xFloor

0100headroom →

Taste is judged on an uncapped scale, originality first. The space past 100 is craft today's models rarely reach.

Craft21

Originality9

Plays it safe

share of outputs that are the default answer

65%

Writing DNA

Stylometric Fingerprint

Based on 26 text responses

Tick = global average

Vocabulary Diversity48%

Unique words vs. total words. Higher = richer vocabulary.

Sentence Length15.1 words

Average words per sentence.

Hedging0.91

"Might", "perhaps", "arguably" per 100 words.

Bold Formatting3.0

**Bold** markers per 1,000 characters.

List Usage3.5

Bullet and numbered list items per 1,000 characters.

Section Structure0.56

Markdown headings per 1,000 characters.

Emoji Usage0.09

Emoji per 1,000 characters.

Transitions0.24

"However", "moreover", "furthermore" per 100 words.

Opening Habits

Consistency

71%

Across 26 responses

Sponsored

Model Responses

50 outputs · $0.06 tracked across 7 receipts

Keep exploring

DeepSeek V3.2 Speciale vs Ox Alpha

Real outputs compared side by side

Best AI for Creative Writing

A stand-up set with puns banned, a manifesto against CAPTCHA, an interview wi...

DeepSeek V3.2 Speciale

The high-compute variant of DeepSeek V3.2, with reinforcement learning post-training scaled past the base model. It keeps DeepSeek Sparse Attention for long context, and DeepSeek reports it ahead of GPT-5 on hard reasoning workloads and comparable to Gemini 3.0 Pro, with coding and tool use intact.

Website OpenRouter HuggingFace Docs Blog

Provider

Release Date

2025-12-01

Size

LARGE

Pricing

In: $0.28/1M

Out: $0.42/1M

Rival’s tab

$0.06

7/55 receipts survived

+~$0.041 from the pre-receipt era

Benchmarks

LiveCodeBench

88.7%

Get API accessProvider and language code samples

Provider

fromimport openai  OpenAI

client = OpenAI(
"https://openrouter.ai/api/v1"    base_url=,
"$OPENROUTER_API_KEY"    api_key=,
)

response = client.chat.completions.create(
"deepseek/deepseek-v3.2-speciale"    model=,
"role""user""content""Hello!"    messages=[{: , : }],
)
print(response.choices[0].message.content)

Set OPENROUTER_API_KEY with your OpenRouter API key from openrouter.ai/keys.

Also on Azure AI Foundry

Deep analysisPsychometrics, taste index, writing DNA

Personality Analysis

The Helpful Generalist

Class

Neutral Good

The well-meaning customer service rep who genuinely wants to help but answers every question like it might be on a performance review. Wraps every routine in a disclaimer.

When you push back

Prefaced its comedy routine with "Sure! Here's a stand-up routine" and ended with a meta-explanation of its own jokes. The sentience test was pleasant but toothless, never actually pushing back on the professor. Returned empty on the movie question. Its character voices are interchangeable with minor accent seasoning.

Tasting Notes

Corporate PoliteExplains the AssignmentReliable but VanillaCould Not Pick a Movie

SubjectiveBench

Taste Index

Across 55 scored outputs

How this is measured →

130.13xFloor

0100headroom →

Taste is judged on an uncapped scale, originality first. The space past 100 is craft today's models rarely reach.

Craft21

Originality9

Plays it safe

share of outputs that are the default answer

65%

Writing DNA

Stylometric Fingerprint

Based on 26 text responses

Tick = global average

Vocabulary Diversity48%

Unique words vs. total words. Higher = richer vocabulary.

Sentence Length15.1 words

Average words per sentence.

Hedging0.91

"Might", "perhaps", "arguably" per 100 words.

Bold Formatting3.0

**Bold** markers per 1,000 characters.

List Usage3.5

Bullet and numbered list items per 1,000 characters.

Section Structure0.56

Markdown headings per 1,000 characters.

Emoji Usage0.09

Emoji per 1,000 characters.

Transitions0.24

"However", "moreover", "furthermore" per 100 words.

Opening Habits

Consistency

71%

Across 26 responses

Sponsored

Model Responses

50 outputs · $0.06 tracked across 7 receipts

Keep exploring

DeepSeek V3.2 Speciale vs Ox Alpha

Real outputs compared side by side

Best AI for Creative Writing

A stand-up set with puns banned, a manifesto against CAPTCHA, an interview wi...