ZMIME
Comparison · 3 models · Updated Nov 22, 2026

DeepSeek V4 Flash Vision Exp vs Qwen3.8 Flash Next vs Mercury 2.5

No independent capability benchmark covers every model here yet. The score uses price and context window.No independent capability benchmark covers every model here yet. The score uses context window.

  1. DeepSeek

    DeepSeek V4 Flash Vision Exp

    Released Aug 21, 2026

    68/100
    • ECI—
    • Price$0.216 / $0.647
    • Context1M
  2. Alibaba (Qwen)

    Qwen3.8 Flash Next

    Released Aug 27, 2026

    65/100
    • ECI—
    • Price$0.20 / $0.50
    • Context262K
  3. Inception

    Mercury 2.5

    Released Sep 8, 2026

    62/100
    • ECI—
    • Price$0.20 / $0.75
    • Context260K
  4. DeepSeek

    DeepSeek V4 Flash Vision Exp

    Released Aug 21, 2026

    64/100
    • ECI—
    • Price$0.216 / $0.647
    • Context1M
  5. Alibaba (Qwen)

    Qwen3.8 Flash Next

    Released Aug 27, 2026

    43/100
    • ECI—
    • Price$0.20 / $0.50
    • Context262K
  6. Inception

    Mercury 2.5

    Released Sep 8, 2026

    43/100
    • ECI—
    • Price$0.20 / $0.75
    • Context260K
01 — Verdict

Capability evidence is incomplete

There is not enough shared capability evidence to name an overall winner. The scores (DeepSeek V4 Flash Vision Exp 68/100, Qwen3.8 Flash Next 65/100, Mercury 2.5 62/100) compare the available specifications and price. The score weighs price 71%, context window 29%. None of the shared benchmarks cover every model here yet, so capability is left out of this verdict.There is not enough shared capability evidence to name an overall winner. The scores (DeepSeek V4 Flash Vision Exp 64/100, Qwen3.8 Flash Next 43/100, Mercury 2.5 43/100) compare the available specifications only. The score weighs context window 100%. None of the shared benchmarks cover every model here yet, so capability is left out of this verdict.

  • CapabilityNot enough dataNo independent benchmark covers every model here yet
  • Lowest priceQwen3.8 Flash NextQwen3.8 Flash Next $0.275 · DeepSeek V4 Flash Vision Exp $0.323 · Mercury 2.5 $0.338 per 1M tokens (3:1 blend)
  • Longest contextDeepSeek V4 Flash Vision ExpDeepSeek V4 Flash Vision Exp 1,000,000 · Qwen3.8 Flash Next 262,144 · Mercury 2.5 260,000 tokens
  • Widest inputsQwen3.8 Flash NextDeepSeek V4 Flash Vision Exp: Text, Images · Qwen3.8 Flash Next: Text, Images, Video · Mercury 2.5: Text
  • Self-hostingDeepSeek V4 Flash Vision Exp and Qwen3.8 Flash NextPublishes downloadable weights (MIT and qwen-community-1.0)
How the score is built
MeasureWeightDeepSeek V4 Flash Vision ExpQwen3.8 Flash NextMercury 2.5
Price71%707369
Context window29%644343
Specs & price only100%68/10065/10062/100

Left out because at least one model lacks the data: capability and inputs & features. The remaining weights were rescaled.

How the score is built
MeasureWeightDeepSeek V4 Flash Vision ExpQwen3.8 Flash NextMercury 2.5
Context window100%644343
Specs only100%64/10043/10043/100

Left out because at least one model lacks the data: capability and inputs & features. The remaining weights were rescaled.

02 — Side by side

Every spec in one table

Highlighted cells lead their row. Unavailable specifications are marked “Not published”. Publisher results include their evaluation settings and source.

DeepSeek V4 Flash Vision Exp vs Qwen3.8 Flash Next vs Mercury 2.5 specifications side by side
SpecificationDeepSeek V4 Flash Vision ExpDeepSeekQwen3.8 Flash NextAlibaba (Qwen)Mercury 2.5Inception
Capability
Capabilities Index (ECI)Not publishedNot publishedNot published
ECI rankNot publishedNot publishedNot published
ECI confidence intervalNot publishedNot publishedNot published
ECI evaluation dateNot publishedNot publishedNot published
Documented capabilitiesExperimental multimodal DeepSeek V4 Flash model for image understanding, coding, and agentic workOpen-weight experimental preview of the Qwen4 architecture: hybrid-attention MoE (125B total, 6B active) with vision encoder for coding, agent tasks, and image and video understandingReasoning model for deliberate analysis, multi-step problem solving, and tool use
Model parametersNot publishedNot publishedNot published
Output modalitiesTextTextText
Publisher-reported benchmarks
Terminal-Bench 2.1max effort83.9scoreReported resultNot publishedNot published
NL2Repomax effort57.7scoreReported resultNot publishedNot published
CyberGymmax effort75.3scoreReported resultNot publishedNot published
DeepSWEmax effort59.3scoreReported resultNot publishedNot published
Toolathlon-Verifiedmax effort75.9scoreReported resultNot publishedNot published
DSBench-Hard · internalmax effort63.6scoreReported resultNot publishedNot published
AutomationBench · publicmax effort25.7scoreReported resultNot publishedNot published
ApexBenchPublisher-reported result36.5pass@1Reported resultNot publishedNot published
Agents' Last ExamPublisher-reported result27.3scoreReported result51.2scoreReported resultNot published
ChartographyPublisher-reported result64.3scoreReported resultNot publishedNot published
ZeroBenchPublisher-reported result35pass@5Reported resultNot publishedNot published
DeepSWE 1.1Publisher-reported resultNot published58.7scoreReported resultNot published
SWE-Bench Pro · Qwen refined and corrected task setPublisher-reported resultNot published62.5scoreReported resultNot published
SWE-Bench MultilingualPublisher-reported resultNot published81scoreReported resultNot published
NL2RepoPublisher-reported resultNot published48.1scoreReported resultNot published
CoWorkBench · internalPublisher-reported resultNot published73.9scoreReported resultNot published
JobBenchPublisher-reported resultNot published55.7scoreReported resultNot published
Agents' Last ExamPublisher-reported resultNot published24.3pass@1Reported resultNot published
Toolathlon-VerifiedPublisher-reported resultNot published73.5pass@1Reported resultNot published
IFBenchPublisher-reported resultNot published81.3scoreReported resultNot published
GPQA DiamondPublisher-reported resultNot published91.7scoreReported resultNot published
Humanity's Last Examwithout tools; GPT-4o judgeNot published35.9scoreReported resultNot published
LiveCodeBench 6Publisher-reported resultNot published91.9scoreReported resultNot published
ClawEval-MMPublisher-reported resultNot published64.4pass@3Reported resultNot published
ClawEval-MMPublisher-reported resultNot published60.4average scoreReported resultNot published
RecreationBench · internalPublisher-reported resultNot published49.9scoreReported resultNot published
AndroidWorldPublisher-reported resultNot published84.5scoreReported resultNot published
OSWorld 2.0Publisher-reported resultNot published19.4binary completion rateReported resultNot published
OSWorld 2.0Publisher-reported resultNot published52.3partial scoreReported resultNot published
Vision2Webgpt-5.4-2026-03-05 judgeNot published64scoreReported resultNot published
ERQAPublisher-reported resultNot published72.3scoreReported resultNot published
LVBenchPublisher-reported resultNot published76.6scoreReported resultNot published
RealWorldQAPublisher-reported resultNot published88.5scoreReported resultNot published
MathVision · corrected annotationswithout code interpreterNot published90.6scoreReported resultNot published
MathVision · corrected annotationswith code interpreterNot published95.7scoreReported resultNot published
CharXiv Reasoningwithout code interpreterNot published84.6scoreReported resultNot published
CharXiv Reasoningwith code interpreterNot published90.6scoreReported resultNot published
Price per million tokens
Input$0.216$0.20 (best)$0.20 (best)
Output$0.647$0.50 (best)$0.75
Cached inputNot publishedNot published$0.02
Blended (3:1)$0.323$0.275 (best)$0.338
Long-context rateSame rateSame rateSame rate
Price sourceMedian of 15 providersMedian of 5 providersOfficial Inception API
Limits
Context window1,000,000 tokens (best)262,144 tokens260,000 tokens
Max inputNot publishedNot publishedNot published
Max output384,000 tokens (best)131,072 tokens65,536 tokens
Inputs and features
TextYesYesYes
ImagesYesYesNo
PDFsNoNoNo
AudioNoNoNo
VideoNoYesNo
ReasoningYesYesYeslow · medium · high
Tool callingYesYesYes
Structured outputNot publishedNot publishedYes
File attachmentsYesYesNo
Temperature controlYesYesYes
Reasoning controlalwaysalwayseffort
Availability
WeightsOpenMITOpenqwen-community-1.0Proprietary
API model IDNot publishedNot publishedmercury-2.5
API providers16 (best)61
ReleasedAug 21, 2026Aug 27, 2026Sep 8, 2026
Knowledge cutoffNot publishedNot publishedNot published
Publisher documentationModel weightsModel weightsPublisher documentation
Source metadata updatedSep 1, 2026Aug 27, 2026Sep 8, 2026
Data sourcesmodels.dev · Epoch AImodels.dev · Epoch AImodels.dev · Epoch AI
Catalog checkedNov 22, 2026Nov 22, 2026Nov 22, 2026
03 — Cost

What would a month cost?

Enter your expected volume in millions of tokens. List prices only; caching and batch discounts would lower these.

  • DeepSeek V4 Flash Vision Exp$3.45
  • Qwen3.8 Flash Next$3.00
  • Mercury 2.5$3.50
04 — Questions

Which should you choose?

Which is better: DeepSeek V4 Flash Vision Exp, Qwen3.8 Flash Next or Mercury 2.5?

There is not enough shared capability evidence to name an overall winner. The scores (DeepSeek V4 Flash Vision Exp 68/100, Qwen3.8 Flash Next 65/100, Mercury 2.5 62/100) compare the available specifications and price. The score weighs price 71%, context window 29%. None of the shared benchmarks cover every model here yet, so capability is left out of this verdict.There is not enough shared capability evidence to name an overall winner. The scores (DeepSeek V4 Flash Vision Exp 64/100, Qwen3.8 Flash Next 43/100, Mercury 2.5 43/100) compare the available specifications only. The score weighs context window 100%. None of the shared benchmarks cover every model here yet, so capability is left out of this verdict.

Which is cheaper, DeepSeek V4 Flash Vision Exp, Qwen3.8 Flash Next or Mercury 2.5?

Qwen3.8 Flash Next is cheaper at $0.20 input / $0.50 output per million tokens (median across 5 API providers). DeepSeek V4 Flash Vision Exp costs $0.216 input / $0.647 output per million tokens (median across 15 API providers); Mercury 2.5 costs $0.20 input / $0.75 output per million tokens (official Inception API price). At a typical mix of three input tokens to one output token, that is $0.275 per million tokens for Qwen3.8 Flash Next versus $0.323 for DeepSeek V4 Flash Vision Exp (1.2× as much) and $0.338 for Mercury 2.5 (1.2× as much).

Which scores higher on benchmarks?

There is no independent benchmark that covers all three models yet. DeepSeek V4 Flash Vision Exp has not been scored yet, Qwen3.8 Flash Next has not been scored yet and Mercury 2.5 has not been scored yet.

Which is better for coding?

There are no published SWE-bench Verified results for DeepSeek V4 Flash Vision Exp, Qwen3.8 Flash Next and Mercury 2.5 yet, so there is no like-for-like coding score. Test both on a sample of your own tasks. All three support tool calling for agent workflows.

Which has the bigger context window?

DeepSeek V4 Flash Vision Exp has the largest context window at 1,000,000 tokens, against 262,144 for Qwen3.8 Flash Next and 260,000 for Mercury 2.5. Maximum output per response: DeepSeek V4 Flash Vision Exp up to 384,000, Qwen3.8 Flash Next up to 131,072, Mercury 2.5 up to 65,536 tokens.

Which can read images, PDFs, audio or video?

DeepSeek V4 Flash Vision Exp accepts text and images; Qwen3.8 Flash Next accepts text, images and video; Mercury 2.5 accepts text. Qwen3.8 Flash Next handles the widest range of inputs.

Are any of these open source?

DeepSeek V4 Flash Vision Exp and Qwen3.8 Flash Next publishes its weights (MIT and qwen-community-1.0) and can be self-hosted; Mercury 2.5 is proprietary.

Which is newer?

Mercury 2.5 is the newest, released Sep 8, 2026. Qwen3.8 Flash Next came out Aug 27, 2026; DeepSeek V4 Flash Vision Exp came out Aug 21, 2026.

How do you decide the winner?

This rating uses price (71%, blended price per million tokens (3 input : 1 output), log scale); context window (29%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 22, 2026.This rating uses context window (100%, maximum tokens per request, log scale). Dimensions missing for any model are dropped and the remaining weights rescaled, so every model is judged on the same evidence. Specs and prices come from public model listings and the labs’ own API pages; capability scores come from independent benchmark runs. Data updated Nov 22, 2026.