TokenMix Research Lab · 2026-04-24
DeepSeek open-sourced V4 on April 24, 2026 — one day after GPT-5.5 shipped — and did it in two variants: V4-Pro (1.6 trillion parameters, 49B active) and V4-Flash (284 billion parameters, 13B active). Both Apache 2.0, both 1M context, both featuring the new CSA + HCA hybrid attention. Pricing is aggressive: V4-Flash at $0.14/$0.28 per MTok (matching V3.2's floor), V4-Pro at
| Claim | Status |
|---|---|
| Both variants released April 24, 2026 | Confirmed |
| V4-Pro: 1.6T total / 49B active (MoE) | Confirmed |
| V4-Flash: 284B total / 13B active (MoE) | Confirmed |
| 1M context on both variants | Confirmed |
| Apache 2.0 license on both | Confirmed |
| Weights on Hugging Face | Confirmed |
| CSA + HCA hybrid attention | Confirmed (official technical report) |
| V4-Pro: .74 input / $3.48 output per MTok | Confirmed (DeepSeek API pricing) |
| V4-Flash: $0.14 / $0.28 per MTok | Confirmed |
| V4 beats Claude Sonnet 4.5 on agentic coding | Confirmed (DeepSeek self-reported) |
| V4-Pro rivals GPT-5.5 on SWE-Bench | Partial — V4-Pro ~85% Verified vs GPT-5.5 88.7%; within 4 points |
| V4-Pro replaces frontier closed models entirely | No — still 10-15% gap on SWE-Bench Pro vs Opus 4.7 |
| DeepSeek filed delay per Huawei chip bottleneck (March reports) | Resolved — shipped April 24 |
| Dimension | V4-Pro | V4-Flash |
|---|---|---|
| Total parameters | 1.6T | 284B |
| Active per token | 49B | 13B |
| Architecture | Sparse MoE + CSA/HCA | Sparse MoE + CSA/HCA |
| Context | 1M | 1M |
| License | Apache 2.0 | Apache 2.0 |
| Input price | .74 / MTok | $0.14 / MTok |
| Output price | $3.48 / MTok | $0.28 / MTok |
| Cache hit input | ~$0.14 / MTok (92% off) | ~$0.03 / MTok (80% off) |
| Throughput | ~60-120 tok/s | ~150-300 tok/s |
| Self-host VRAM (FP8) | ~800GB+ | ~150GB |
| Realistic self-host | 8× H200 or 16× H100 | 2× H100 or 4× A100 |
| Best workload | Frontier reasoning, complex agents | High-throughput production, batch jobs |
The 3.8× parameter gap (49B vs 13B active) and 12× price gap ( .74 vs $0.14) are the big decision drivers. Flash is economically disruptive — close to V3.2 pricing with a clearly better architecture. Pro is a legitimate frontier competitor at fractional cost.
DeepSeek V4's attention mechanism is the architectural novelty worth understanding:
CSA (Compressed Sparse Attention) — handles local context with aggressive compression, suitable for the "near past" tokens where fine-grained detail matters most.
HCA (Heavily Compressed Attention) — handles the 1M-token long tail with even heavier compression, where rough semantic recall is what you need, not token-level fidelity.
The hybrid routes tokens through whichever mechanism fits the distance from the current generation position. Net effect: 1M context without the O(n²) memory blowup that makes most long-context claims fall apart in practice.
Early third-party testing shows V4 maintains solid recall up to ~700K tokens, with noticeable degradation only past 850K. Compare to Claude Opus 4.7's 1M (stable to ~900K) and GPT-5.5's 256K cap — V4 is competitive on long context, not best-in-class.
List prices are only half the story. Cache-hit prices are where DeepSeek genuinely disrupts:
| Tier | Standard Input | Cache-Hit Input | Output |
|---|---|---|---|
| V4-Pro | .74 / MTok | ~$0.14 / MTok (92% off) | $3.48 / MTok |
| V4-Flash | $0.14 / MTok | ~$0.03 / MTok (80% off) | $0.28 / MTok |
For comparison:
The takeaway: On workloads with stable system prompts (agent systems, RAG with fixed contexts, long document Q&A), V4-Flash at $0.03/MTok cache hit is 17× cheaper than GPT-5.5's cache hit and 17× cheaper than Opus 4.7's.
For a team running 10M input tokens/day with 70% cache hit:
| Benchmark | V4-Pro | V4-Flash | Context |
|---|---|---|---|
| SWE-Bench Verified | ~85% | ~78% | vs GPT-5.5 88.7%, Opus 4.7 87.6% |
| SWE-Bench Pro | ~55% | ~48% | vs GPT-5.5 58.6%, Opus 4.7 64.3% |
| AIME 2025 | ~94 | ~88 | vs Step 3.5 Flash 97.3 |
| MMLU | ~89% | ~84% | vs GPT-5.5 92.4% |
| Long-context recall (500K) | Strong | Strong | 1M context supports both |
| Terminal-Bench 2.0 | ~75% | ~68% | vs GPT-5.5 82.7%, Qwen 3.6-27B 59.3% |
The honest read:
Pick V4-Pro when:
Pick V4-Flash when:
Skip both when:
Both variants are Apache 2.0 open-weight, available on Hugging Face. Self-hosting is viable:
V4-Pro self-host reality:
V4-Flash self-host reality:
For most teams, hosted API is the right call unless there's a specific reason (compliance, sovereignty, extreme volume). The $0.14/MTok hosted price is hard to beat on TCO for anything short of 100M+ daily tokens.
DeepSeek kept V4-Flash's pricing identical to V3.2 ($0.14/$0.28 per MTok). This is strategic:
For existing V3.2 users, the migration decision is trivial: switch to V4-Flash today, get better model at same price.
Ranking current frontier-tier models on open/closed and price:
V4 slots in at two levels: Pro is the best open-weight 1.6T you can get with Apache 2.0; Flash is the price-performance floor for agentic production.
See our Chinese AI models comparison guide for the broader landscape including Kimi K2.6, Step 3.5 Flash, Qwen 3.6, GLM-5.1, and others.
Q: Which DeepSeek V4 should I use for coding agents? A: V4-Pro if you want quality parity with Claude/GPT at 1/10 the cost. V4-Flash if you need fast iteration or are running high-throughput jobs where quality gap is acceptable.
Q: How does V4-Pro compare to GPT-5.5? A: V4-Pro is ~4 points behind GPT-5.5 on SWE-Bench Verified, ~3 points behind on MMLU, but costs 1/3 of GPT-5.5 and is open-weight. For most workloads, V4-Pro is the better economic pick; for absolute frontier coding on SWE-Bench Pro, Claude Opus 4.7 still leads.
Q: Can I self-host V4-Pro on a single H100? A: No. V4-Pro requires ~800GB VRAM in FP8 — you need 8-16 GPUs. V4-Flash at 284B fits in 2× H100, which is realistic for small teams.
Q: Does V4 support OpenAI-compatible API?
A: Yes. Both variants are accessible via api.deepseek.com/v1 with the same OpenAI-compatible endpoint as V3.2. Drop-in migration.
Q: Is V4's 1M context real or marketing? A: Third-party testing shows stable recall up to ~700K, noticeable degradation past 850K. For production, stay under ~600K for critical accuracy. This is comparable to Claude Opus 4.7's 1M behavior.
Q: What's the cache-hit economics on V4? A: V4-Pro cache hits drop input from .74 to ~$0.14/MTok (92% off). V4-Flash cache hits drop from $0.14 to ~$0.03/MTok (80% off). For agent workloads with stable system prompts, real effective input cost is ~$0.03-0.14/MTok.
Q: When will DeepSeek V4.1 or V4.2 release? A: Based on DeepSeek's historical cadence, expect a V4.1 minor update within 3-4 months. Full V5 likely Q2 2027.
By TokenMix Research Lab · Updated 2026-04-24