DeepSeek V3.1 model profile
Evidence snapshot:
DeepSeek V3.1 is a DeepSeek model profile with a 163,840-token context window. The dated catalog snapshot records text modalities and interfaces for prompt caching, reasoning controls, structured output, text generation, tool calling.
DeepSeek-V3.1 is a large hybrid reasoning model (671B parameters, 37B active) that supports both thinking and non-thinking modes via prompt templates. It extends the DeepSeek-V3 base with a two-phase long-context training process, reaching up to 128K tokens, and uses FP8 microscaling for efficient inference. Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs The model improves tool use, code generation, and reasoning efficiency, achieving performance comparable to DeepSeek-R1 on difficult benchmarks while responding more quickly. It supports structured tool calling, code agents, and search agents, making it suitable for research, coding, and agentic workflows. It succeeds the DeepSeek V3-0324 model and performs well on a variety of tasks.
DeepSeek V3.1 is a reference profile: DeepSeek publishes it, Kendr documents it for comparison, and it carries no Kendr alias to call.
DeepSeek V3.1 is an independent knowledge profile. It does not claim that Kendr hosts the model, quote a Kendr customer price, promise API availability, or advertise a Kendr routing receipt.
DeepSeek V3.1 accepts text and returns text. The snapshot records prompt caching, reasoning controls, structured output, text generation, and tool calling as capabilities or interfaces.
The dated reference snapshot lists $0.25 input and $0.95 output per million tokens. Cached input is $0.13 per million tokens.
The 2026-09-06 snapshot includes 7 Design Arena categories and AutoExacto results for 9 provider observations. Publisher-reported launch results are labeled and are not presented as Kendr measurements; scores from different suites are not treated as interchangeable.
DeepSeek V3.1 ranked #100 in the cited trailing-7-day third-party catalog usage dataset as of 2026-09-06. Observed within the cited trailing-7-day third-party catalog usage dataset; this is not global AI market share.
Profile data was reviewed for the 2026-09-06 snapshot. Claims retain their source dates, and unavailable fields are shown as unavailable instead of estimated.
Profile facts
- Profile type
- Model research profile
- Provider or publisher
- DeepSeek
- Context window
- 163,840 tokens
- Snapshot date
- 2026-09-06
- Knowledge cutoff
- 2025-03-31
- Reference catalog ID
- deepseek/deepseek-chat-v3.1
When to pick DeepSeek V3.1
DeepSeek V3.1 is a reference profile: DeepSeek publishes it, Kendr documents it for comparison, and it carries no Kendr alias to call.
Each condition below is derived from this model's own values in the 2026-09-06 snapshot, compared only against models publishing the same field.
- Reach for it when the model has to call tools. Tool calling is recorded in the snapshot, so this route can drive an agent loop rather than only answer in prose.
- Look elsewhere when you need to call it through Kendr today. This is a research reference profile with no Kendr alias. The facts and evidence here are published for comparison; the model is not routable on Kendr.
- Look elsewhere when the prompt is long. The 164K-token context window is in the bottom quarter of the 163 models publishing a limit. Long documents need chunking or a wider-context route.
Context, modalities, and identity
DeepSeek V3.1 accepts text and returns text. The snapshot records prompt caching, reasoning controls, structured output, text generation, and tool calling as capabilities or interfaces.
- Provider or publisher
- DeepSeek
- Context window
- 163,840 tokens
- Maximum output
- 144,900 tokens
- Input modalities
- text
- Output modalities
- text
- Knowledge cutoff
- 2025-03-31
Model overview
DeepSeek-V3.1 is a large hybrid reasoning model (671B parameters, 37B active) that supports both thinking and non-thinking modes via prompt templates. It extends the DeepSeek-V3 base with a two-phase long-context training process, reaching up to 128K tokens, and uses FP8 microscaling for efficient inference. Users can control the reasoning behaviour with the reasoning enabled boolean. Learn more in our docs The model improves tool use, code generation, and reasoning efficiency, achieving performance comparable to DeepSeek-R1 on difficult benchmarks while responding more quickly. It supports structured tool calling, code agents, and search agents, making it suitable for research, coding, and agentic workflows. It succeeds the DeepSeek V3-0324 model and performs well on a variety of tasks.
- Reference model ID
- deepseek/deepseek-chat-v3.1
- Canonical version
- deepseek/deepseek-chat-v3.1
- Hugging Face ID
- deepseek-ai/DeepSeek-V3.1
Capabilities and supported controls
Capabilities and parameters are reported from the dated catalog and source records; their presence does not guarantee identical behavior across every provider route.
- Supported parameters
- frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, top_p
- prompt caching
- reasoning controls
- structured output
- text generation
- tool calling
Dated pricing snapshot
The dated reference snapshot lists $0.25 input and $0.95 output per million tokens. Cached input is $0.13 per million tokens.
These are dated third-party reference prices, not Kendr prices or an availability offer. Routes, tiers, caching, region, and provider policy can change the landed price.
- Price date
- 2026-09-06
- Input
- $0.25 per 1M tokens
- Cached input
- $0.13 per 1M tokens
- Output
- $0.95 per 1M tokens
Provider routes, performance, and uptime
The snapshot retains 8 provider endpoints. Provider prices, context limits, p50 performance, and uptime can differ by route and are not Kendr guarantees.
- Active provider endpoints
- 5
- Best p50 latency
- 0.58 s
- Best p50 throughput
- 40 tok/s
- Availability with routing
- 99.92% over the sampled window
- Availability without routing
- 91.72% over the sampled window
- Performance date
- 2026-09-06
| Provider | Quantization | Input / 1M | Output / 1M | Cache read / 1M | Context | p50 latency | p50 throughput | Uptime (1d) |
|---|---|---|---|---|---|---|---|---|
| DeepInfra | fp4 | $0.25 / 1M | $0.95 / 1M | $0.13 / 1M | 163,840 tokens | 0.83 s | 10 tok/s | 99.99% |
| SiliconFlow | fp8 | $0.27 / 1M | $1 / 1M | No verified rate | 163,840 tokens | 1.27 s | 17 tok/s | 96.96% |
| NovitaAI | fp8 | $0.27 / 1M | $1 / 1M | $0.135 / 1M | 131,072 tokens | 1.73 s | 23 tok/s | 99.96% |
| AtlasCloud | fp8 | $0.3 / 1M | $0.95 / 1M | $0.13 / 1M | 131,072 tokens | 1.71 s | 38 tok/s | 99.73% |
| CoreWeave | fp8 | $0.55 / 1M | $1.65 / 1M | $0.55 / 1M | 161,000 tokens | 0.58 s | 40 tok/s | 100.00% |
| MARA | unknown | $0.6 / 1M | $1.7 / 1M | No verified rate | 131,072 tokens | 1.66 s | 31 tok/s | 95.24% |
| SambaNova | fp8 | $0.65 / 1M | $1.5 / 1M | No verified rate | 131,072 tokens | 1.77 s | 34 tok/s | 98.56% |
| Google Vertex | unknown | $0.6 / 1M | $1.7 / 1M | No verified rate | 163,840 tokens | No recent p50 | No recent p50 | 0.00% |
Benchmark evidence and limitations
The 2026-09-06 snapshot includes 7 Design Arena categories and AutoExacto results for 9 provider observations. Publisher-reported launch results are labeled and are not presented as Kendr measurements; scores from different suites are not treated as interchangeable.
Third-party catalog fields normalized from the cited model dataset; benchmark methodology and coverage differ by source.
- AutoExacto coverage
- 9 provider observations over 32 days
| Design Arena category | Elo | Win rate | Rank |
|---|---|---|---|
| 3d | 1114 | 47.9% | #83 |
| codecategories | 1128 | 47.8% | #85 |
| dataviz | 1115 | 46.2% | #87 |
| gamedev | 1113 | 47.1% | #84 |
| svg | 1002 | 38.2% | #82 |
| uicomponent | 1107 | 47.1% | #84 |
| website | 1135 | 48% | #85 |
AutoExacto provider benchmarks
Rolling provider observations from the cited third-party benchmark view. The lookback is 32 days; GPQA Diamond and TAU-Bench Airline measure different abilities and should not be blended into one score.
| Provider | GPQA Diamond | TAU-Bench Airline | Runs |
|---|---|---|---|
| AtlasCloud | 80.19% | — | 4 |
| auto-routing | 74.54% | 64.1% | 4 |
| CoreWeave | 78.65% | 41.35% | 4 |
| DeepInfra | 76.55% | 40.23% | 4 |
| Google Vertex | 76.12% | 47.2% | 3 |
| MARA | 75.33% | — | 3 |
| NovitaAI | 81.74% | 73.44% | 4 |
| SambaNova | 69.48% | — | 2 |
| SiliconFlow | 0.06% | 52.78% | 4 |
Popularity and market context
DeepSeek V3.1 ranked #100 in the cited trailing-7-day third-party catalog usage dataset as of 2026-09-06. Observed within the cited trailing-7-day third-party catalog usage dataset; this is not global AI market share.
- Third-party catalog rank
- #100
- Global market share
- Not inferred
- Observation date
- 2026-09-06
Frequently asked questions
When should I use DeepSeek V3.1?
DeepSeek V3.1 is a reference profile: DeepSeek publishes it, Kendr documents it for comparison, and it carries no Kendr alias to call. Reach for it when the model has to call tools: Tool calling is recorded in the snapshot, so this route can drive an agent loop rather than only answer in prose. Look elsewhere when you need to call it through Kendr today: This is a research reference profile with no Kendr alias. The facts and evidence here are published for comparison; the model is not routable on Kendr. Look elsewhere when the prompt is long: The 164K-token context window is in the bottom quarter of the 163 models publishing a limit. Long documents need chunking or a wider-context route.
What is DeepSeek V3.1?
DeepSeek V3.1 is a DeepSeek model profile with a 163,840-token context window. The dated catalog snapshot records text modalities and interfaces for prompt caching, reasoning controls, structured output, text generation, tool calling.
What context window does DeepSeek V3.1 have?
The dated profile lists 163,840 tokens of context and up to 144,900 output tokens. Provider routes, variants, and runtime configuration can impose lower effective limits.
What benchmark evidence is available for DeepSeek V3.1?
The 2026-09-06 snapshot includes 7 Design Arena categories and AutoExacto results for 9 provider observations. Publisher-reported launch results are labeled and are not presented as Kendr measurements; scores from different suites are not treated as interchangeable. Third-party catalog fields normalized from the cited model dataset; benchmark methodology and coverage differ by source.
Is DeepSeek V3.1 available through Kendr?
This is an independent knowledge profile, not a Kendr-hosted availability claim. Check Kendr's live public model API for currently enabled Kendr aliases.
Does the popularity rank represent DeepSeek V3.1's global market share?
DeepSeek V3.1 ranked #100 in the cited trailing-7-day third-party catalog usage dataset as of 2026-09-06. Observed within the cited trailing-7-day third-party catalog usage dataset; this is not global AI market share. No global market-share percentage is inferred when the source does not publish one.
Sources and evidence dates
- DeepSeek V3.1 third-party catalog record (catalog, checked 2026-09-06)
- Third-party model catalog methodology (methodology, checked 2026-09-06)
- DeepSeek official model documentation (primary, checked 2026-09-06)
- Design Arena leaderboard and methodology (benchmark, checked 2026-09-06)
- DeepSeek V3.1 third-party catalog record (benchmark, checked 2026-09-06)