Models / DeepSeek/ DeepSeek V3.1 Terminus

DeepSeek V3.1 Terminus

Name: DeepSeek V3.1 Terminus
Author: DeepSeek

DeepSeek · released Sep 22, 2025 · deepseek-ai/DeepSeek-V3.1-Terminus

InputOutput

Type: Open weightsMIT License
Params: 685B
Context: 164K

37B active per word · about 123K words of context

Our take

Written Aug 3, 2026

DeepSeek V3.1 Terminus is a large downloadable text model with 685 billion total parameters and 37 billion active per token, released under a permissive MIT licence. It scores notably higher on coding than on general text tasks, and comes with ten hosted options that trade off cost against speed.

Who should pick it

Pick this for research or commercial work that needs a permissive open licence, or for production where you can choose your cost-throughput trade-off. Use SambaNova if latency matters most, Atlas Cloud for balance, or DeepInfra if budget is tighter. Skip it if you need image, audio or video input, or if you need strong maths performance.

The case for it

Extremely large total parameter count with moderate active parameters per token: 685 billion total, 37 billion active.
Strong coding performance relative to its general text score, 46 points higher on the coding leaderboard.
Ten distinct hosted offers spanning a wide range of cost and throughput combinations.
Permissive MIT licence allows commercial use, modification and redistribution.

The case against it

Maths and instruction following lag behind coding and hard-prompt performance, with a 67-point gap between maths and coding scores.
Text-to-text only; no image, audio or video input supported.
Throughput data is sparse and inconsistent, with only six of ten offers listing it and some hosts showing conflicting figures.

How good is it?

What has been measured, on which board, and where DeepSeek V3.1 Terminus placed.

IntelligencePuzzles, maths, exam questions

2.5 of 5

Arena Text (overall)65th of 143 · 1415.1

Arena Hard Prompts 71st of 143Arena Maths 82nd of 139

Also on this board: 1417.2 via Thinking (Aug 2, 2026). Read the pair, not the higher one.

CodingWriting and fixing code on its own

2.5 of 5

Arena Coding79th of 143 · 1439

Arena Coding is the only board that has scored it for this.

Also on this board: 1463.3 via Thinking (Aug 2, 2026). Read the pair, not the higher one.

AgenticPlanning, calling tools, staying on task

not measured

Nobody we watch has scored DeepSeek V3.1 Terminus for this. We would take the rating from Arena Agent (IPS).

WritingWe do not rate this

Scored, not ratedThe placings are on the right.

Two boards come close and neither tests writing: Arena Creative Writing asks people which of two replies they prefer, and LiveBench Language tests whether a model understood a passage. So we show where DeepSeek V3.1 Terminus placed and give it no mark out of five.

Arena Creative Writing 45th of 143 · 1406.8

Also scored, on boards we give no mark for

Arena Instruction Following 73rd of 143

These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done, which is why they get no rating.

Every published score for this model6 scoresEvery figure we hold, from 6 boards, with who ran it and a link to the source — including the boards no rating above is built on.

Arena Codingcoding

1439independentsource ↗

Arena Creative Writingchat

1406.8independentsource ↗

Arena Hard Promptsreasoning

1423.5independentsource ↗

Arena Instruction Followingchat

1393.4independentsource ↗

Arena Mathsmath

1396.1independentsource ↗

Arena Text (overall)chat

1415.1independentsource ↗

Can you run it yourself?

Fits fully in memory on 0 of the 71 devices we track, at Q4.

Fits in memory: weights load entirely on the card
Spills to system RAM: some weights offload; much slower
Too large: will not load even with offload
est: size is calculated; the verdict could change by 10%

A card many people ownToo large

GeForce RTX 4090 · 24 GB

Weights at Q4_K_M431.6 / 24 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Too large for this card. The weights do not fit even with part of them offloaded to system memory.

One step upToo large

Apple M1 Pro (16-core GPU) · 32 GB

Weights at Q4_K_M431.6 / 32 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Too large for this card. The weights do not fit even with part of them offloaded to system memory.

One step downToo large

Radeon RX 7900 XT · 20 GB

Weights at Q4_K_M431.6 / 20 GBest

Usable contextNot calculated for spilled setups

Decode speedNot estimated for spilled setups

Too large for this card. The weights do not fit even with part of them offloaded to system memory.

Your hardware

Checking your profile…

Memory use by level

Against a 24 GB card.

Q4_K_M

recommended

431.6 GBest

Too large

Q5_K_M

506.3 GBest

Too large

Q8_0

756.4 GBest

Too large

All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.

Check against your own machine → · All 71 devices, with every size →

Or rent it from someone else

10 providers, from 13 listings.

Cheapest published offer

OpenRouter

Cheapest of 13 live listings. Picked at the widest standard context we hold, within one quantisation slice, so the numbers beside it are a price one host actually charges.

per 1M tokens: $0.25 in / $0.95 out
Context served: 164K
Throughput: Not measured

Current provider offers with price, context and prompt-privacy answers
Provider	In / out per 1M tokens	Context	Throughput	Trains on prompts	Logs prompts	Zero retention
DeepInfrafp4	$0.25 / $0.95	164K	8 tok/s	No	No	Confirmed
AtlasCloudfp8	$0.30 / $0.95	131K	32 tok/s	No	Yesunknown period	Unknown
OpenRouter	$0.25 / $0.95	164K	not measured	Unknown	Unknown	Unknown
DeepInfrafp4	$0.25 / $0.95	164K	not measured	Unknown	Unknown	Unknown
SiliconFlowfp8	$0.27 / $1.00	164K	13 tok/s	No	No	Confirmed
Novita AI	$0.27 / $1.00	131K	not measured	Unknown	Unknown	Unknown
Novita AIfp8	$0.27 / $1.00	131K	26 tok/s	No	No	Confirmed
StreamLake	$0.34 / $1.03	128K	58 tok/s	No	Yesunknown period	Unknown
SambaNovafp8	$0.65 / $1.50	131K	54 tok/s	No	No	Confirmed
CoreWeavefp8	$0.55 / $1.65	161K	50 tok/s	No	No	Confirmed
Google Vertex AIus-west2	$0.60 / $1.70	164K	28 tok/s	No	No	Confirmed
Mara	$0.60 / $1.70	131K	28 tok/s	No	No	Confirmed
SambaNova	$3.00 / $4.50	131K	not measured	Unknown	Unknown	Unknown

Across the 13 listings we hold: 9 say they do not train on prompts, 0 say they do and 4 do not say. 7 appear in the zero-retention registry we check; the rest are unknown to us rather than confirmed either way.

What each host's API supports

From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.

✓
Supported: ✗
Not supported
Not published: host gave no parameter list

API features per host
Provider	Tool calling	JSON output	Strict schema
DeepInfrafp4	✓	✓	✓
AtlasCloudfp8	✗	✓	✓
OpenRouter	✓	✓	✓
DeepInfrafp4
SiliconFlowfp8	✓	✓	✓
Novita AI
Novita AIfp8	✓	✓	✓
StreamLake	✗	✗	✓
SambaNovafp8	✓	✗	✗
CoreWeavefp8	✓	✓	✓
Google Vertex AIus-west2	✓	✓	✓
Mara	✗	✓	✓
SambaNova

Tool calling: 7 of 13 listings say yes, 3 say no, 3 publish no parameter list. JSON output: 8 of 13 listings say yes, 2 say no, 3 publish no parameter list. Strict schema: 9 of 13 listings say yes, 1 says no, 3 publish no parameter list.

Models people weigh against DeepSeek V3.1 Terminus

Shortlist any of these to compare them side by side.

Weighed against itGLM 5.2 Kimi K2.6 GLM 5.1

Compare them field by field →

When we formed this view

And whether anything has moved since.

Dates behind this page

Aug 3, 2026Price changeDeepInfra raised DeepSeek V3.1 Terminus pricing by 8%input +8% ($0.25 → $0.27 per 1M tokens)detected

Aug 3, 2026Price changeDeepInfra cut DeepSeek V3.1 Terminus pricing by 7%input −7% ($0.27 → $0.25 per 1M tokens)detected

Aug 3, 2026Price changeOpenRouter cut DeepSeek V3.1 pricing by 7%input −7% ($0.27 → $0.25 per 1M tokens); output −5% ($1.00 → $0.95 per 1M tokens); cache read −4% ($0.14 → $0.13 per 1M tokens)detected

Aug 3, 2026Price changeOpenRouter raised DeepSeek V3.1 Terminus pricing by 8%input +8% ($0.25 → $0.27 per 1M tokens); output +5% ($0.95 → $1.00 per 1M tokens); cache read +4% ($0.13 → $0.14 per 1M tokens)detected

Aug 2, 2026Price changeDeepInfra raised DeepSeek V3.1 Terminus pricing by 8%input +8% ($0.25 → $0.27 per 1M tokens)detected

Aug 2, 2026Price changeOpenRouter cut DeepSeek V3.1 pricing by 7%input −7% ($0.27 → $0.25 per 1M tokens); output −5% ($1.00 → $0.95 per 1M tokens); cache read −4% ($0.14 → $0.13 per 1M tokens)detected

Aug 2, 2026Price changeOpenRouter raised DeepSeek V3.1 Terminus pricing by 8%input +8% ($0.25 → $0.27 per 1M tokens); output +5% ($0.95 → $1.00 per 1M tokens); cache read +4% ($0.13 → $0.14 per 1M tokens)detected

Aug 2, 2026BenchmarkScored 1439 on Arena Codingleaderboardannounced

Aug 2, 2026BenchmarkScored 1406.8 on Arena Creative Writingleaderboardannounced

Aug 2, 2026BenchmarkScored 1423.5 on Arena Hard Promptsleaderboardannounced

Prices last checked 4h ago

What we do not know about this model yet

We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
3 of 13 listings publish no parameter list, so what their API accepts is unknown to us.
Nothing we hold says whether an endpoint streams, so we do not show it either way.
4 of 13 listings do not say whether they train on prompts.

Licence and identifiers

What the licence allowsMIT License, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.

Licence

MIT License

permissiveCommercial use allowed

Fully permissive: do anything with attribution. No patent grant, unlike Apache-2.0.

Identifiers

Hugging Face: deepseek-ai/DeepSeek-V3.1-Terminus
Architecture: Mixture of experts
Modality record: text->text
Catalogue slug: deepseek-deepseek-v3-1-terminus

Machine-readable model card (omc.json) →

Explore furtherWhere to run DeepSeek V3.1 Terminus Best open-weight models

Something wrong on this page? Tell us