Throughput (tokens / sec)
median 80.2tokens/s
High-throughput agentic work with 1M context
Model deprecated
DeepSeek V4 Flash is deprecated but remains available through our API. Its retirement date will be announced soon. We recommend migrating to an active model before its retirement. Visit our model library to find a suitable replacement.
Run instantly. Pay only for what you use.
DeepSeek V4 Flash is DeepSeek's efficiency-focused open-weight model: a 284B-parameter MoE that activates just 13B per token, with a 1M-token context window and MIT-licensed weights. Three thinking levels — non-think, think-high, and think-max — let you pick speed or reasoning depth per request.
Its hybrid sparse attention is built for long context: at 1M tokens it runs on roughly 27% of the compute and 10% of the KV cache of DeepSeek V3.2, so whole-repo prompts and heavily cached agent context stay fast and affordable.
The result is production-grade agentic coding at a small-model price: 79% on SWE-bench Verified and 91.6 on LiveCodeBench in max-thinking mode, with strict function calling and JSON output available in every mode.
1Mtoken context window
Sparse attention runs 1M-token prompts on ~27% of the compute of DeepSeek V3.2.
13Bactive of 284B total
MoE routing keeps per-request cost and latency near small-model levels.
97%avg. cache hit rate on Entrim
Measured on coding-agent workloads — repeated context bills at the cache-read rate.
DeepSeek V4 Flash is strongest where agents reason over a lot of context: coding loops, tool chains, and document pipelines that need long memory without dense-model cost.
Holds a plan across long tool-call chains — the 0731 re-post-train pushes Terminal-Bench 2.1 to 82.7, up from 61.8 on the preview build.
Non-think, think-high, and think-max — the 0731 re-post-train sharpens the reasoning modes hardest (DeepSWE 7.3 → 54.4), so you pay for depth only where a request needs it.
Hybrid sparse attention reads 1M tokens with ~10% of the KV cache of V3.2 — whole repos and document sets in one prompt.
Strict function calling and JSON mode in every thinking level — tool arguments and typed pipelines that parse the first time.
Long-running coding agents
Plans, edits, and verifies across long tool-call chains — 79% on SWE-bench Verified in max-thinking mode.
Whole-repo and long-document analysis
The 1M-token window holds a full repository or document set in one prompt, and sparse attention keeps it affordable.
High-volume, cost-sensitive fleets
13B active parameters and deep cache-read discounts keep per-request cost flat as traffic scales.
Structured extraction to JSON
Strict function calling and JSON mode return schema-valid output for pipelines — in every thinking mode.
Image, audio, or video inputs
DeepSeek V4 Flash is text-only. Route visual inputs to DeepSeek V4.1 Flash or Gemma 4 31B.
Fact recall without retrieval
Like other small-activation models it is weak on standalone knowledge recall (34.1% on SimpleQA-Verified) — pair it with RAG for factual workloads.
Tight output-token budgets in think-max
Max thinking emits long reasoning traces that drive up output tokens and latency. Use non-think or think-high for cost- and latency-sensitive replies.
Compare Entrim’s pricing, powered by an optimized inference runtime, against other providers using the same token counts per request.
median 80.2tokens/s
median 1,051ms
82.7Terminal-Bench 2.1
Agentic coding and shell tasks run end-to-end in a real terminal.
54.4DeepSWE
Resolves real software-engineering issues across long tool-call chains.
70.3Toolathlon (verified)
Sustains long multi-tool workflows without losing the plan.
Code examples
curl -X POST "https://api.entrim.ai/v1/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $ENTRIM_API_KEY" \ -d '{ "model": "deepseek-ai/DeepSeek-V4-Flash", "messages": [ { "role": "user", "content": "Why do they call it a building if it is already built?" } ] }'SDKs & docs
Use familiar SDK patterns with quickstart examples for common production setups.
OpenAI-compatible
Keep the request format your team already knows.
No cold starts
Production requests are served without model spin-up delays.
Agentic coding with image input and 1M context
Dense multimodal coding with thinking control
Multimodal long-context work with thinking control