Qwen3.8-27B vs DeepSeek V4 for agentic coding: what the numbers show

Updated 13 November 2026

Qwen and DeepSeek both publish agentic coding scores, so this looks easier than the Sonnet question. It isn't, and the reason is in the tables themselves: each vendor lists Opus 4.6 Max as a reference, and they give it different scores. Here's what each model card says and why the rows can't be merged into one ranking.

What each card reports

TestQwen3.8-27B (Qwen's card)Opus 4.6 Max (Qwen's card)DeepSeek V4-Pro Max (DeepSeek's card)Opus 4.6 Max (DeepSeek's card)
SWE-bench Pro61.753.455.457.3
Terminal Bench73.0 (version 2.1)78.2 (version 2.1)67.9 (version 2.0)65.4 (version 2.0)
LiveCodeBench90.3 (v6)88.8 (v6)93.588.8
SWE-bench VerifiedNot reportedNot reported80.680.8

Why the rows don't line up

SWE-bench Pro as each vendor reports it. Both cards list Opus 4.6 Max, and they disagree by 3.9 points, so the rows can't be read as one ranking.

On SWE-bench Pro, Qwen's card puts Opus 4.6 Max at 53.4 and DeepSeek's card puts it at 57.3. It's the same model, so the 3.9-point gap comes from setup. Qwen says it corrected problem tasks in the benchmark and re-ran the other models, and used Opus's officially reported score for that row. DeepSeek's card gives no harness notes. Terminal Bench differs too: Qwen reports version 2.1 and DeepSeek version 2.0.

Read across the two cards and Qwen3.8-27B looks ahead on SWE-bench Pro (61.7 against 55.4). Read each card against its own Opus row and the picture is different: Qwen is 8.3 points above Opus on its card, and DeepSeek is 1.9 points below Opus on its own. Neither comparison is wrong, and neither tells you how the two models would do on the same tasks.

How the models differ

Qwen3.8-27BDeepSeek V4
Size27 billion parametersV4-Flash has 284 billion in total, with 13 billion active per token
LicenceApache 2.0MIT
Context window256K tokens natively1M tokens (V4-Flash)
WeightsOpenOpen
Where you can run itHosted on NeverMetered, or on your own hardwareDeepSeek's API, other hosts, or your own much larger hardware

What they cost to use

DeepSeek's API is a cheap per-token option. Off-peak, at the calculator's default workload, a typical day of about 8.5M tokens comes to roughly $8 a month on V4.1 Flash and $34 on V4 Pro, and a busy day of about 86.7M comes to $86 and $340. Peak hours cost double. Those are list-price estimates with an assumed cache hit rate. A flat plan only pays off when your usage is heavy enough: see flat rate versus cheap per-token APIs and the cost calculator.

How to choose

  • If you need a 1M-token context window, or you want the model with the most published coding results, look at DeepSeek V4.
  • If you want a flat monthly price and several agents running at once, Qwen3.8-27B on a plan priced by parallel streams is the fit. See how parallel streams work.
  • If you can't decide, run the same five real tasks through both and compare. The test in the Sonnet comparison works for any pair of models.
Is Qwen3.8-27B better than DeepSeek V4 for coding?
The two vendors' tables can't answer that. They measure on different benchmark versions and different setups, and disagree about the same Opus model by almost four points on SWE-bench Pro. Test both on your own tasks.
Which is cheaper?
DeepSeek's API is cheap per token, so light use costs very little. A flat plan wins when you run agents for many hours a day. The cost calculator shows where the break-even sits for your usage.
Do you offer DeepSeek?
No. NeverMetered serves one model, Qwen3.8-27B. DeepSeek's own open-source coding agent can run on it, though: see the DeepSeek Harness setup.
Sources