Skip to main content
AI Choice Engine

Model comparison

DeepSeek V3.1 vs DeepSeek V4 Flash

Two DeepSeek tiers compared on the figures that decide which one a workload actually needs.

DeepSeek V3.1: profile checked October 25, 2026DeepSeek V4 Flash: profile checked October 25, 2026Individual provider fields may changeEvidence confidence: Insufficient data — see receipts below

DeepSeek

DeepSeek V3.1

Budget · Open weights

vs

DeepSeek

DeepSeek V4 Flash

Balanced · Open weights

AI model capability comparison
SpecificationDeepSeek V3.1DeepSeek V4 Flash
ProviderDeepSeekDeepSeek
TierBudgetBalanced
Context window128KWinner: 1M
Max outputNot verifiedUnverifiedNot verifiedUnverified
Input / 1M tokensNot verifiedUnverifiedNot verifiedUnverified
Output / 1M tokensNot verifiedUnverifiedNot verifiedUnverified
WeightsOpenOpen
Parameters671B total / 37B active (MoE)284B total / 13B active (MoE)
Reasoning levelsNot verifiedUnverifiedlow, high, max
Modalitiestexttext
LicenseMITMIT
API model idNot verifiedUnverifieddeepseek-v4-flash
ReleasedAugust 21, 2025April 24, 2026
SWE-bench Verified · source date DeepSeek V4 Flash: 2026-04-24Not verifiedUnverified79
GPQA Diamond · source date DeepSeek V4 Flash: 2026-04-24Not verifiedUnverified88.1
Humanity's Last Exam · source date DeepSeek V4 Flash: 2026-04-24Not verifiedUnverified45.1
MMLU-Pro · source date DeepSeek V4 Flash: 2026-04-24Not verifiedUnverified86.2

Dates beside benchmark scores refer to the linked source’s publication or capture date unless that source explicitly states when it measured the result. A shared URL alone does not establish comparable runs; unverified setups are marked. Prices are USD per million tokens at standard rates, excluding batch and caching discounts. Values we could not confirm from the provider are shown as “Not verified” rather than estimated.

Benchmark receipts

Where each score comes from, and how far it can be compared across models.

  • Source date2026-04-24
    DeepSeek family HF model card (vendor, Think Max, exact match)

    Broad knowledge across 14 domains with harder, reasoning-focused multiple-choice questions.

    Comparable with caveatFrontier models cluster in the high 80s–90s; near-ceiling differences are weak evidence. CoT vs direct answering must match.

  • Source date2026-04-24
    DeepSeek family HF model card (vendor, Think Max, with tools; 34.8 no tools)

    Frontier-knowledge ceiling: extremely hard multi-domain questions written to be near-impossible without deep expertise.

    Comparable with caveatSubset/tool configurations (text-only vs with-tools) differ between vendors and materially change scores; must match configuration to compare.

  • Source date2026-04-24
    DeepSeek family HF model card (vendor, Think Max, pass@1)

    Graduate-level science reasoning: multiple-choice questions written by domain PhDs.

    Comparable with caveatDo not treat a fixed two-point band as universal measurement uncertainty; it depends on run count and evaluation setup. Chain-of-thought, prompt, and tool policy must match. The small pool limits fine-grained conclusions, and frontier scores can be near ceiling; Artificial Analysis dropped it from its Intelligence Index in v4.3.

  • Source date2026-04-24
    DeepSeek original family HF model card (vendor, Think Max; historical April checkpoint, not the 0731 refresh)

    Real GitHub issue resolution: does the model's patch pass the hidden tests.

    Comparable with caveatPost-audit vendor claims and pre-audit scores sit on different task trust levels; scaffolding (agent harness, compute budget) also dominates results. Never aggregate across scaffolds.

Pricing tiers

DeepSeek V3.1: No current first-party API or verified third-party tariff. deepseek-chat moved to V3.1-Terminus on September 22, 2025; that alias later retired July 24, 2026. MIT weights remain for self-hosting or separately priced hosts. The original checkpoint's hard output cap and selectable hosted effort menu remain unverified. Current Flash/Pro API prices do not price the original V3.1 checkpoint.

DeepSeek V4 Flash: Retired 2026-09-10 — no longer sold as the original checkpoint. Historical checkpoint tariffs are not verified here. The legacy deepseek-v4-flash ID now routes to V4.1-Flash-1017 at $0.30/$1.20 peak or $0.15/$0.60 off-peak, with $0.006/$0.003 cache reads. MIT weights remain. The 384K current API output ceiling and vision support belong to the routed revision, not these original text-only weights.

BudgetOpen weightsRecord checked October 25, 2026

DeepSeek V3.1

DeepSeek V3.1 is a legacy open-weight (MIT) 671B/37B MoE with a 128K context — superseded by V3.1-Terminus, V3.2 and then V4.

Best for

  • Cost-sensitive hosted agents
  • High-volume coding assist
  • Open-weight deployments

Watch out

Legacy: DeepSeek's first-party alias moved to V3.1-Terminus on September 22, 2025. Pin the exact V3.1 checkpoint; current hosted rates depend on provider. The original card supports thinking/non-thinking templates, not a verified selectable low/high/max menu. Self-hosting memory depends on quantization, offloading and hardware. The current first-party pricing table lists Flash-1017 and Pro-0813 only; absence of a third-party tariff here means unknown, not free.

BalancedOpen weightsRecord checked October 25, 2026

DeepSeek V4 Flash

DeepSeek V4 Flash was the efficient open-weight (MIT) sibling of V4 Pro — retired from DeepSeek's API on 2026-09-10 and replaced by V4.1-Flash.

Best for

  • Open-weight deployments
  • Cost-sensitive inference
  • High-volume serving

Watch out

The original V4 Flash API checkpoint retired September 10; its old ID is still accepted and now serves Flash-1017. Original MIT weights remain available for self-hosting. Original API low/high/max settings are historical controls, not evidence of a matching native hosted menu or successor performance.

When the cheaper one wins

One of these rows is missing a verified output price, so this page will not name a cheaper winner. Run the model picker against the actual job instead of guessing.

Run the model picker

Evidence confidence: Insufficient data

How strong and complete the evidence behind this comparison is — not a prediction of which model is better. Hard limitations apply: missing pricing; evidence too sparse.

  • Pricing incomplete — No verified token pricing for DeepSeek V3.1 and DeepSeek V4 Flash.
  • 3/5 core specs verified on both sides — Not published for at least one side: max output, reasoning levels.
  • No directly aligned benchmark setup — No benchmark has matching cited sources, source dates, units, effort and recorded run details on both models.
  • Verified within the last 90 days — Older of the two catalog checks was 1 day ago.
  • Both models carry source citations — Each side has at least two catalog sources on record.

Source receipts

Each catalog figure was checked against the provider or an independent second source on the date shown.

Related comparisons

Diving deeper on one model? DeepSeek V3.1 · DeepSeek V4 Flash

Common questions

DeepSeek V3.1 vs DeepSeek V4 Flash

Answered from the verified figures on this page rather than general guidance.

Which has the larger context window, DeepSeek V3.1 or DeepSeek V4 Flash?
DeepSeek V4 Flash accepts 1M tokens against 128K for DeepSeek V3.1. This only matters if you routinely send very long documents or large codebases.
Should I use DeepSeek V3.1 or DeepSeek V4 Flash?
We classify DeepSeek V3.1 in the budget tier and DeepSeek V4 Flash in the balanced tier. Test whether either meets your accuracy, tool-use and latency requirements before choosing on price.

Next step

Choosing between them

The verified figures that separate this pair, computed from the catalog rather than restated boilerplate.

  • Context: DeepSeek V4 Flash takes 1M against 128K for DeepSeek V3.1 — only decisive if your prompts approach the smaller window.
  • Positioning: DeepSeek V3.1 sits in the budget tier, DeepSeek V4 Flash in the balanced tier. Compare accuracy, reliability, latency, and cost on representative tasks before choosing.
Run the model pickerEstimate stack costAll model comparisons

Choosing a harness rather than a model — Cursor, Copilot, Claude Code, Windsurf, or Lovable? Compare agentic harnesses · Latest releases.