DeepSeek V4
Open-weight frontier intelligence at a radical price — MIT-licensed, $0.435/$0.87 per million tokens after a permanent 75% cut, with LiveCodeBench 93.5%, SWE-bench Verified 80.6%, 1M context, and an unmatched 384K output window. Released April 24, 2026.
Model Specs
- Released
- Apr 2026
- Context window
- 1.0M tokens
- Max output
- 384K tokens
- Capabilities
- reasoningfunction-callinglong-contextcost-efficient
- Modalities
- text
About this model
DeepSeek V4 is the model that re-proved the open-weight frontier. Released April 24, 2026 as a two-model family — **V4-Pro** (the 1.6-trillion-parameter MoE flagship with 49B active parameters) and **V4-Flash** (a 284B/13B workhorse) — both ship under the **MIT license** with weights on Hugging Face, free for commercial use and self-hosting. On the Artificial Analysis Intelligence Index, V4-Pro scores 52, the #2 open-weights model, a ten-point generational jump over V3.2.
The coding numbers compete with models ten times the price: **93.5% on LiveCodeBench (the highest published), 80.6% on SWE-bench Verified** (within 0.2 of Opus 4.6-class results), 55.4% on SWE-Bench Pro, and a 3206 Codeforces Elo. The architecture is the quiet revolution: a hybrid of Compressed Sparse Attention and Heavily Compressed Attention that needs only ~27% of V3.2's compute per token at 1M context and 10% of the KV cache — which is how DeepSeek affords a **1M-token context** and an industry-unique **384K-token output window** (3x larger than any Western flagship) at these prices.
And the prices are the story: after DeepSeek made its launch discount **permanent** in May 2026, V4-Pro costs $0.435/M input and $0.87/M output — roughly **6-11x cheaper than GPT-5.5 and 23-57x cheaper than Fable 5** — with cache hits at $0.004/M. The trade-offs are real: it's slower (44.8 t/s), very verbose in its reasoning, and its hallucination rate on AA-Omniscience is high (94%) — it guesses rather than abstains. Reach for DeepSeek V4 when cost-per-intelligence is the deciding factor: high-volume coding, batch reasoning, self-hosted deployments, and massive-output tasks like full-codebase generation.
Key Strengths
Radical price-to-intelligence ratio
$0.435/M input, $0.87/M output — permanently. Intelligence Index 52 at roughly 6-11x less than GPT-5.5 and a fraction of any Western flagship. Running AA's entire benchmark suite costs ~$268 vs $4,800+ for premium rivals.
Highest published LiveCodeBench score
93.5% on contamination-resistant competitive coding — above every Western flagship — plus SWE-bench Verified 80.6% and a 3206 Codeforces Elo.
MIT open weights
Full weights on Hugging Face under MIT — self-host, fine-tune, and deploy commercially with no restrictions. The strongest permissively-licensed model family available.
384K output window — industry's largest
Generate entire codebases, books, or datasets in one response. Western flagships cap at 128K output; V4 triples that with its Think Max mode.
Novel efficiency architecture
Hybrid CSA/HCA sparse attention: ~27% of V3.2's per-token compute at 1M context and 10% of the KV cache — the architectural breakthrough that funds the pricing.
V4-Flash sibling for volume
The 284B/13B Flash variant runs at $0.14/$0.28 with Intelligence Index 47 and 87 t/s — frontier-adjacent quality at prices that round to zero.
How it compares
DeepSeek V4 is the price-performance outlier of the frontier tier — compare on cost, openness, and reliability needs.
| vs. Model | Verdict | Outcome |
|---|---|---|
| Qwen3.7 Max | The two Chinese frontier leaders: DeepSeek V4 is ~6x cheaper, MIT open-weights, and edges LiveCodeBench (93.5 vs 91.6); Qwen3.7 Max is smarter overall (Index 57 vs 52), 4x faster (171 vs 45 t/s), and far more precise on instructions (IFBench 80.5 vs 71.3). Budget and self-hosting → DeepSeek; production agents → Qwen. | Depends |
| Gemini 3.5 Flash | Gemini 3.5 Flash is smarter (Index 55 vs 52), 3x faster, multimodal, and far more reliable on facts — at ~3x the price ($1.50/$9 vs $0.435/$0.87). DeepSeek counters with open weights, the 384K output window, and stronger competitive coding. Interactive and multimodal → Gemini; volume text/code → DeepSeek. | Other wins |
| GPT-5 Mini | Similar budget positioning, different classes: DeepSeek V4 is a frontier model (Index 52, SWE-V 80.6%) at budget prices, while GPT-5 Mini (Index 41) is a true lightweight. V4 wins nearly every capability metric; GPT-5 Mini wins on speed-to-first-token and OpenAI ecosystem fit. For quality per dollar, V4 is the stronger pick. | Wins most cases |
Pros
- Cheapest frontier-class API: $0.435/$0.87 per M tokens (permanent)
- MIT open weights — self-host, fine-tune, commercial use
- LiveCodeBench 93.5% — highest published score
- SWE-bench Verified 80.6%, Codeforces 3206 Elo
- 384K output window — 3x larger than Western flagships
- 1M context via novel sparse-attention architecture
- V4-Flash sibling at $0.14/$0.28 for extreme volume
Things to consider
- High hallucination rate (94% on unanswerable questions) — guesses rather than abstains
- Slow output for its class (44.8 t/s) and very verbose reasoning
- Trails Western flagships on composite intelligence (52 vs 60-65)
- Text-only — no image, video, or audio input
- Knowledge cutoff unconfirmed by official documentation
- Weaker on agentic terminal tasks (Terminal-Bench Hard 41.7%)
Best use cases
High-volume coding
The LiveCodeBench leader at ~2% of flagship prices — code generation, review, and refactoring pipelines where unit economics decide feasibility.
Self-hosted deployments
MIT weights mean full control: on-prem inference for regulated data, fine-tuning on proprietary corpora, no per-token vendor lock-in.
Massive-output generation
384K output tokens in one response — full project scaffolds, long-form documentation, dataset synthesis that other models must chunk.
Batch reasoning at scale
Classification, extraction, and analysis over millions of items where a 6-11x price difference compounds into entire budgets.
Long-context analysis on a budget
1M context with AA-LCR 65% at cents per request — contract sets, log archives, codebases.
Math & competitive programming
Codeforces 3206 Elo and strong competition-math scores make it a powerful training partner and solution checker.
How to use it on Renas AI
- 1
Step 1
Open AI Chat on Renas
Navigate to AI Chat in the Renas dashboard and pick the most capable DeepSeek model available to your plan.
- 2
Step 2
Pick the thinking depth
V4 offers Non-Think, Think High, and Think Max modes. Use Non-Think for fast extraction, Think Max for hard problems — the 384K output window opens at maximum depth.
- 3
Step 3
Verify factual claims
V4 guesses rather than abstains (94% hallucination rate on unanswerable questions). For factual work, ask for sources or pair it with a low-hallucination model as verifier — its strength is reasoning and code, not recall honesty.
- 4
Step 4
Exploit the economics
At these prices, generate multiple candidate solutions and pick the best — sampling strategies that would be prohibitive on flagships cost cents here.
Pricing
Pricing on Renas AI
Pay-as-you-go credits, no API keys, no rate limits.
~2,000,000 words on a 10,000-credit Spark plan
Frequently asked questions
Other text models on Renas AI
Open-weight frontier on Renas
Access leading DeepSeek models with your Renas AI subscription credits — no API key, no setup, no per-seat fees.
Try DeepSeek on Renas