models.fyiPreview data

DeepSeek-V3.2

DeepSeekOpen weights

The price-performance anchor of the market: 90%+ of frontier quality at ~5% of frontier cost on the official API.

Index @ Medium

69.5

Context

128K

$/M in · out

$0.28 · $0.42

Median speed

60 tok/s

Released

Dec 9, 2025

02·Which effort for which task

The sweet spot, per task

Cheapest effort tier that keeps ≥96% of this model's peak quality on the benchmarks behind each task type. Below the floor number, quality is not okay — it just fails quietly.

Everyday writing & Q&A

Medium

Chat, drafting, summaries, general questions.

keeps 98.0% of peak−63% cost

Code assist

High

Autocomplete, single functions, small fixes.

keeps 100.0% of peakpeak tier needed

Coding agents

High

Real repos, multi-file changes, end-to-end tickets.

keeps 100.0% of peakpeak tier needed

Math & quantitative

High

Derivations, statistics, financial models.

keeps 100.0% of peakpeak tier needed

⚠ at Low: only 70% of peak — avoid

Hard research problems

High

Novel, frontier, genuinely difficult questions.

keeps 100.0% of peakpeak tier needed

⚠ at Low: only 78% of peak — avoid

Agents & tool use

High

Long workflows with tools, policies, and state.

keeps 100.0% of peakpeak tier needed

03·The effort ladder

Quality and cost, tier by tier

On everyday tasks, Medium keeps 98.0% of High quality at $0.0015 vs $0.0039 per task.

Intelligence Index by effort

Cost per typical task by effort

04·Benchmark detail

Scores at Medium

Low Medium High

at Mediumat High (peak)

Benchmark	Medium	Peak	What it predicts
MMLU-Pro	82.9	85.0	Everyday Q&A · Drafting emails & docs · Summarizing articles
GPQA Diamond	76.1	83.0	Technical analysis · Root-cause debugging · Scientific writing
Humanity's Last Exam	20.2	22.0	Novel research questions · Frontier problem solving
AIME 2025	75.5	90.0	Financial modeling · Statistics · Algorithm design
SWE-bench Verified	66.8	72.0	Coding agents · Bug fixing in real repos · Multi-file refactors
LiveCodeBench	73.3	79.0	Writing functions from scratch · Algorithms · Code autocomplete
τ²-bench	71.5	78.0	Tool-using agents · Customer-facing automation · Multi-step workflows
IFEval	86.7	88.0	Structured output (JSON) · Templated generation · Batch pipelines

05·Host variants

Same weights, different model

Hosts serve DeepSeek-V3.2 at different quantizations and stacks. Index below is measured per host at Medium effort — the headline number only applies to the reference serving.

Host	Quant	Index	Δ vs reference	$/M in · out	tok/s	TTFT	Uptime	Gateway id
DeepSeek (official)reference	fp8	69.5	—	$0.28 · $0.42	60	0.90s	98.9%	deepseek-v3.2
Fireworks	fp8	69.1	-0.5	$0.56 · $1.68	180	0.32s	99.2%	deepseek-v3.2@fireworks
Together AI	fp8	69.2	-0.3	$0.60 · $1.70	150	0.40s	99.0%	deepseek-v3.2@together
DeepInfra	int8	67.5	-2.1	$0.30 · $0.88	110	0.50s	98.4%	deepseek-v3.2@deepinfra

06·Drift

Is it still the same model this week?

DeepSeek (official)FireworksTogether AIDeepInfra

Weekly Intelligence Index · dashed rules mark serving-change events

Feb 15, 2026DeepSeek (official)-0.7 ptsrecovered Mar 1, 2026

Chat-endpoint prompt-template bug degraded instruction following for two weeks.

07·Call it

One key, this model, your effort

The gateway speaks the OpenAI Chat Completions dialect. Pick a host with deepseek-v3.2@host syntax, set reasoning_effort, and the sweet spot above becomes one line of config. Full docs →

chat with DeepSeek-V3.2

curl https://models.fyi/api/v1/chat/completions \
  -H "Authorization: Bearer $MFYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v3.2",
    "reasoning_effort": "medium",
    "messages": [{"role": "user", "content": "Hello"}]
  }'
# local dev: replace https://models.fyi with http://localhost:3000