DeepSeek-V4.1-Flash
A multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively.
DeepSeek’s own description of the model, from DeepSeek-V4.1-Flash model card. Served here as deepseek-ai/deepseek-v4.1-flash.
On this roster: 1M context with reasoning on by default, for agent runs that carry long histories and want the reasoning channel exposed.
What DeepSeek built it for
in the vendor’s words- This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially improving cost efficiency for input-heavy agentic workloads.
- Combined with FP4 main KV caching (E2M1 format, one E4M3 scale per 16 channels), these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash.
- The model supports a continuously controllable reasoning effort setting (integer 1–100) that trades inference cost for accuracy.
Quoted from DeepSeek-V4.1-Flash model card, retrieved 13 Sept 2026. Thinking is on by default on this endpoint (the model template's own default; the gateway injects no reasoning default of its own) and reasoning_effort is passed through when a request sets it. The card's benchmark results all use the maximum effort.
Published benchmarks
DeepSeek’s figures, not ours| Benchmark | What it measures | Score | How DeepSeek ran it |
|---|---|---|---|
| GPQA Diamond | Graduate-level science questions | 90.9 | The card's preamble to this table: all instruct results use the maximum effort setting (reasoning_effort=100); evaluations use temperature=1.0, top_p=0.95. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| HLE 39.1 on the text-only subset (the card's † footnote) | Reasoning, full set | 36.8 | The card's preamble to this table: all instruct results use the maximum effort setting (reasoning_effort=100); evaluations use temperature=1.0, top_p=0.95. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| Codeforces | Competitive programming rating | 3471 | The card's preamble to this table: all instruct results use the maximum effort setting (reasoning_effort=100); evaluations use temperature=1.0, top_p=0.95. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| MathArena Apex | Competition mathematics | 65.6 | The card's preamble to this table: all instruct results use the maximum effort setting (reasoning_effort=100); evaluations use temperature=1.0, top_p=0.95. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| Terminal-Bench 2.1 | Agentic terminal coding | 90.6 | The card: code agent benchmarks are evaluated with the Minimal mode of DeepSeek Harness and a 1M-token context window; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| Terminal-Bench 3.0 | Agentic terminal coding | 30.0 | The card: code agent benchmarks are evaluated with the Minimal mode of DeepSeek Harness and a 1M-token context window; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| Terminal-Bench 4.0 | Agentic terminal coding | 31.2 | The card: code agent benchmarks are evaluated with the Minimal mode of DeepSeek Harness and a 1M-token context window; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| DeepSWE v1.1 | Agentic coding, real-world software engineering | 74.2 | The card: the mini-SWE harness is used for DeepSWE v1.1 to align with its official setup; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| ProgramBench | Agentic programming | 20.3 | The card: code agent benchmarks are evaluated with the Minimal mode of DeepSeek Harness and a 1M-token context window; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| NL2Repo-Bench | Repository generation from natural language | 64.0 | The card: code agent benchmarks are evaluated with the Minimal mode of DeepSeek Harness and a 1M-token context window; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| CyberGym | Security, vulnerability reproduction | 88.1 | The card: all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100 (no per-benchmark harness note for this row). Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| SEC-Bench Pro | Security engineering | 62.8 | The card: the Claude Code harness is used for SEC-Bench Pro to align with its official setup; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| ExploitGym | Security, exploit development | 15.3 | The card: all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100 (no per-benchmark harness note for this row). Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| HLE w/ tools | Reasoning with tools | 63.9 | The card: all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100 (no per-benchmark harness note for this row). Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| AutomationBench | Workflow automation | 54.8 | The card: Agent's Last Exam and AutomationBench use their official scaffolds; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| Agent's Last Exam | Agentic reasoning | 31.8 | The card: Agent's Last Exam and AutomationBench use their official scaffolds; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| Chartography w/ tools | Visual agent, chart reading | 78.9 | The card: visual agent benchmarks use the Claude Code harness with a 512k-token context window; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| BabyVision w/ tools | Visual agent, visual reasoning | 89.6 | The card: visual agent benchmarks use the Claude Code harness with a 512k-token context window; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
| ZeroBench-main w/ tools | Visual agent, Pass@5 | 49.0 | The card: visual agent benchmarks use the Claude Code harness with a 512k-token context window; all agentic evaluations use temperature=1.0, top_p=0.95 at reasoning_effort=100. Mode: Max reasoning effort (reasoning_effort=100), per the card's table heading. |
Every score above is DeepSeek’s own published result for this model, copied from DeepSeek-V4.1-Flash model card on 13 Sept 2026. We did not run these benchmarks and we do not restate them as our own. Each vendor ran its own harness with its own judge, so these tables are not a league table: scores from two different cards are not comparable, even where the benchmark name matches. Each model here shows only its own maker’s table, for that reason.
Recommended usage
what DeepSeek recommendsSampling
For recommended sampling parameters - the default path, and what we serve, DeepSeek recommends temperature=1.0, top_p=0.95. That is what this endpoint sends when your request omits them, so the default request shape is the vendor’s recommendation rather than ours. An explicit value in your request always wins.
DeepSeek publishes a second set for the card's alternative top_p: temperature=1.0, top_p=1.0. We send top_p 0.95, the value every evaluation on the card used; set top_p 1.0 explicitly if you want the alternative.
Thinking
Thinking is on by default and the reasoning streams before the answer: in reasoning_content on Chat Completions, as a reasoning item on Responses and as a thinking block on Messages. reasoning_effort is passed through when a request sets it (low, high and max are accepted); the card's own benchmark results use the maximum. Reasoning tokens bill as output.
Context length
DeepSeek gives the native window as 1,048,576 tokens, which is what we serve: the full checkpoint window, with input and output sharing it. Prompt and output share the window: the served registry caps a prompt at 786,432 tokens and the output at 262,144 tokens (the card recommends max_tokens of 256K or more), and a request that sets no max_tokens gets 32,768. Reasoning tokens count toward the output and bill as output - send a generous max_tokens and stream.
What to point it at
DeepSeek names these workloads for this model:
- “substantially improving cost efficiency for input-heavy agentic workloads”
- “support for contexts of up to one million tokens”
- “large-scale automated synthesis of agent tasks and environments with progressive scaling of data, tasks, and rollouts”
Recommendations on this page are DeepSeek’s, from DeepSeek-V4.1-Flash model card, retrieved 13 Sept 2026. We default to them rather than to a house preference; where our default would differ, this page says so.
What it does on this endpoint
our measurements · not vendor figuresThese figures are ours, measured on the live endpoint: observations, not service guarantees, and not to be confused with the vendor scores above. The protocol, the vantage, the build each figure is pinned to and how to check it against the system_fingerprint in your own response are all on the measurements page. Run your own and use those instead.
Price
per million tokens · no request fee · no subscription| Input | Cached input | Output |
|---|---|---|
| $0.30 | $0.006 | $1.20 |
One price, no tiers. The cached rate applies only to tokens your responses report as cached_tokens — see the prompt caching card below for where that stands on this model. Full price sheet and worked examples.
Capabilities
deepseek-ai/deepseek-v4.1-flashThree API formats, one base URL
OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages, all served from https://api.tiyuvta.ai/v1. Point an existing client at it and set the model id; there is no product-specific SDK.
Tools and structured output
tools and tool_choice, JSON mode, and JSON schema via response_format. The tool-call round trip is part of the release gate on every deploy. This model reasons before it answers and the schema is enforced after that channel closes, so max_tokens has to cover the reasoning and the JSON together. Run out while it is still reasoning and the request fails closed with a named error; run out after the JSON has started and you get the ordinary max_tokens truncation, so budget generously or stream.
Vision
Images go in as image_url content parts and bill as ordinary input tokens; vision costs nothing extra.
Prompt caching
Prompt caching restores retained prefixes and bills only the tokens reported as cached_tokens. In qualification, a 32.7k-token prompt (32,674-32,758 tokens) reported 32,512 cached tokens on 9 of 9 repeats. Cold prompts still need processing. Residency is bounded and shared; a later request can return cold.
Gated before it serves
Every build of this model is reconciled against the vendor’s published model and passes a live gate on all three wire formats plus a tool-call round trip before it takes traffic; a failing gate blocks the release. The correctness gates.
Where it fits best
use it for- Long-context agent loops: the window we serve is the window we publish, and prompts near the top of it answered with every planted needle found in qualification
- Reasoning-heavy work where you want to see the reasoning: thinking is on by default, the channel is exposed on every request shape, and reasoning tokens are billed as output - send a generous max_tokens and stream
- Tool-calling agents on the default tool_choice: native tool calls round-trip and stream as tool_calls deltas, and response_format json_schema is honoured
- Input-heavy agentic work: 552B backbone parameters with 8B active per token during prefill and 16B during decode, so long prompts are the cheap part
Also on the same key and balance: GLM-5.3-Flash, Ornith-1.5 35B-A3B, Qwen3.8 27B, Qwen3-Embedding-8B, Qwen3-Reranker-8B.
First request
deepseek-ai/deepseek-v4.1-flashcurl https://api.tiyuvta.ai/v1/chat/completions \
-H "Authorization: Bearer $TIYUVTA_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "deepseek-ai/deepseek-v4.1-flash",
"messages": [{"role": "user", "content": "Hello"}]}'The default request reasons before it answers, at DeepSeek’s recommended sampling: best quality for hard problems.
- Vendor sources
- huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash, MIT. DeepSeek trained this model; we serve it at its full native context. Benchmark scores and usage recommendations on this page come from DeepSeek-V4.1-Flash model card, retrieved 13 Sept 2026.
- Machine-readable metadata
GET https://api.tiyuvta.ai/v1/modelsfor supported parameters, and the provider feed atGET https://api.tiyuvta.ai/models?schema=openrouter.- Rate limits
- Per account, for abuse control, raised on request; capacity scales with demand rather than sitting behind a published ceiling.