Model comparison

DeepSeek vs Qwen: benchmarks, speed and price

DeepSeek V4.1 Flash is the coding pick (Max-report DeepSWE: 74.2% vs 58.7%); Qwen3.8 Flash adds video and halves Novita AI's input price ($0.15 vs $0.30 per 1M tokens).

Updated · 6 sources · Methodology

Reported Max vs Flash-Next scores, Artificial Analysis (AA) references and Novita AI prices
Dimension Winner Margin Evidence
Coding benchmark (reported) DeepSeek V4.1 Flash (Max) +15.5 percentage points (DeepSWE) DeepSWE v1.1: DeepSeek (Max) 74.2% vs Qwen Flash-Next 58.7%12
Reasoning: GPQA (reported) Qwen3.8-Flash-Next +0.8 percentage points GPQA Diamond: Qwen Flash-Next 91.7% vs DeepSeek (Max) 90.9%12
Reasoning: HLE (reported) DeepSeek V4.1 Flash (Max) +0.9 percentage points HLE: DeepSeek (Max) 36.8% vs Qwen Flash-Next 35.9%12
Price (input / output) Qwen3.8 Flash 50% lower input price; about 61% lower output price Qwen $0.15 / $0.47 vs DeepSeek $0.30 / $1.20 per 1M tokens34
Output speed (AA reference) DeepSeek V4.1 Flash (Max) ~3.6× the AA output rate DeepSeek (Max) 186 vs Qwen Flash-Next 52 tokens/s5

Benchmarks

Official benchmark results

DeepSWE v1.1, NL2Repo-Bench, GPQA Diamond and HLE from the published reports; higher scores are better.

Official reports: DeepSeek V4.1 Flash (Max) vs Qwen3.8-Flash-Next

Swipe to see all columns

Metric DeepSeek V4.1 Flash (Max) Qwen3.8-Flash-Next
GPQA Diamond (reasoning) 90.9% · vendor-reported, 2026 91.7% · vendor-reported, 2026
HLE (reasoning, without tools) 36.8% · vendor-reported, 2026 35.9% · vendor-reported, 2026
DeepSWE v1.1 (coding) 74.2% · vendor-reported, 2026 58.7% · vendor-reported, 2026
NL2Repo-Bench (coding) 64.0 · vendor-reported, 2026 48.1 · vendor-reported, 2026

Sources: DeepSeek-V4.1-Flash1, Qwen3.8-Flash-Next2 · Checked 2026-10-14

Model versions and evaluation settings.

Independent AA model evaluations

Swipe to see all columns

Metric DeepSeek V4.1 Flash (Max) Qwen3.8-Flash-Next
AA Intelligence Index v4.3.2 39 40
AutomationBench-AA 69% 56%
Terminal-Bench 4.0 27% 25%

Source: Artificial Analysis5 · Checked 2026-10-14

Speed and latency

Performance

Artificial Analysis reference models: DeepSeek V4.1 Flash (Max) and Qwen3.8-Flash-Next.

Swipe to see all columns

Metric DeepSeek V4.1 Flash (Max) Qwen3.8-Flash-Next
Output speed (tokens/s) 186 52
First-chunk median (s) 0.96 2.02
Time to first answer token (s) 11.68 40.60

Source: Artificial Analysis5 · Checked 2026-10-14

Specs and price

Pricing

Both models run as serverless endpoints on Novita AI. Prices are per 1M tokens.

Swipe to see all columns

Metric DeepSeek V4.1 Flash Qwen3.8 Flash
Model ID deepseek/deepseek-v4.1-flash qwen/qwen3.8-flash
Input $0.30 $0.15
Output $1.20 $0.47
Cache read $0.006 $0.016
Context window (tokens) 1M 977K
Max output (tokens) 384K 128K
Input types Text, image Text, image, video

Sources: Novita AI DeepSeek V4.1 Flash3, Novita AI Qwen3.8 Flash4 · Checked 2026-10-14

Estimated token cost: 100M input + 20M output tokens, no cache
DeepSeek V4.1 Flash $54.00
Qwen3.8 Flash $24.40
Lower estimated token cost is better.

Which to choose

DeepSeek for coding agents and long outputs, Qwen for cost and video

Choose DeepSeek V4.1 Flash if…

  • Coding agents are the priority: DeepSeek's Max report scores 74.2% on DeepSWE v1.1 vs 58.7% for Flash-Next, with NL2Repo-Bench at 64.0 vs 48.1.
  • A request needs a larger token budget: Novita AI lists 1M context and 384K maximum output, versus 977K and 128K for Qwen.
  • Input-heavy prompts are reused: at equal cache-hit rates above 93.75%, DeepSeek's input-only cost is lower. Output remains more expensive, so this does not establish a lower total cost.
  • Reference-model throughput informs your shortlist: AA measures DeepSeek (Max) at 186 tokens/s vs 52 for Qwen Flash-Next, separate from Novita AI's endpoint specifications.
  • Trade-off: uncached input and output cost $0.30 / $1.20 per 1M tokens, and video input is unavailable.

Choose Qwen3.8 Flash if…

  • Uncached token cost is the priority: an estimated $24.40 vs $54.00 for 100M input and 20M output tokens.
  • Long-document prompts are rarely cached: a 977K context window at $0.15 per 1M input tokens.
  • Your inputs include video; DeepSeek V4.1 Flash takes text and image only.
  • Reasoning workloads also need lower token prices: Flash-Next reports 91.7% on GPQA vs 90.9% for DeepSeek (Max); HLE's point estimates favor DeepSeek, 36.8% vs 35.9%.
  • Trade-off: Flash-Next's reported DeepSWE score is lower (58.7% vs 74.2% for DeepSeek Max), and Novita AI caps the Flash API's output at 128K tokens.

Try both

Call both models with one Novita AI key

Same client, same messages: only the model ID changes, and billing follows the model.

import os

from openai import OpenAI



client = OpenAI(

    base_url="https://api.novita.ai/openai",

    api_key=os.environ["NOVITA_API_KEY"],

)



for model in ["deepseek/deepseek-v4.1-flash", "qwen/qwen3.8-flash"]:

    reply = client.chat.completions.create(

        model=model,

        messages=[{"role": "user", "content": "Explain what this function does."}],

        max_tokens=1024,

    )

    print(model, reply.choices[0].message.content)

Source: Novita AI Chat Completions API6 · Checked 2026-10-14

Methodology

Methodology and sources

Official scores cover DeepSeek V4.1 Flash at reasoning_effort=100 and Qwen3.8-Flash-Next, the open-weight basis of the Flash API. Both DeepSWE reports use temperature=1.0 and top_p=0.95; DeepSeek uses mini-SWE-agent with 1M context, while Qwen selects the highest of two harness results, achieved with mini-SWE-agent at 256K. NL2Repo harnesses differ. Qwen uses an automated model judge for HLE; DeepSeek's model card does not identify the same judge. Score differences are reported point estimates, not a matched endpoint comparison or proof of a practical reasoning advantage.

Novita AI publishes this page and serves both models. Its public model pages supply prices, cache rates and limits, including DeepSeek's fp8 quantization3. Novita states that high is the default DeepSeek reasoning setting. The official DeepSeek scores use Max, rather than that default. AA's DeepSeek (Max) and Flash-Next figures are model references, representing first-party APIs or provider medians rather than a Novita-specific endpoint comparison; first-answer latency includes thinking. The estimated chart cost uses uncached synchronous rates. The 93.75% crossover applies only to input cost at equal cache-hit rate h: $0.30 × (1 − h) + $0.006 × h = $0.15 × (1 − h) + $0.016 × h. No private benchmark was run.

  1. DeepSeek · DeepSeek-V4.1-Flash on Hugging Face huggingface.co · Checked 2026-10-14
  2. Qwen · Qwen3.8-Flash-Next on Hugging Face huggingface.co · Checked 2026-10-14
  3. Novita AI · DeepSeek V4.1 Flash novita.ai · Checked 2026-10-14
  4. Novita AI · Qwen3.8 Flash novita.ai · Checked 2026-10-14
  5. Artificial Analysis · DeepSeek V4.1 Flash vs Qwen3.8 Flash-Next artificialanalysis.ai · Checked 2026-10-14
  6. Novita AI · Chat Completions API docs.novita.ai · Checked 2026-10-14

Run DeepSeek V4.1 Flash and Qwen3.8 Flash on one API

Create an API key and use the two model IDs through Novita AI's OpenAI-compatible API.

Start Building Explore Models

Everything you need to build production AI.

200+ models, on-demand GPUs, and secure agent runtimes — unified under one API. Free to start, scales as you grow.