Что такое DeepSeek V4 GA? Таймлайн от preview до general availability
После трёх месяцев preview DeepSeek V4 GA (general availability) вышел 20 июля 2026. Относительно апрельского preview GA апгрейдит Agent performance, math reasoning и code generation — и впервые вводит peak-valley pricing: сигнал перехода DeepSeek от почти бесплатного доступа к дисциплинированной коммерческой эксплуатации.
| Дата | Веха |
|---|---|
| 24 апреля 2026 | V4 preview + MIT open source: V4-Pro (1,6T) и V4-Flash (284B) |
| Май 2026 | Production-tuned V4-Flash и V4-Pro; API generally available |
| Июнь 2026 | Output price V4-Pro −75% навсегда (до $0,87/M tokens) |
| 29 июня 2026 | Email всем API users: GA mid-July, первое раскрытие peak-valley billing |
| 19 июля 2026 | Selected devs получают GA gray-release; медиа: «full version imminent» |
| 20 июля 2026 | GA release (дата публикации статьи) |
| 24 июля 2026 | Legacy deepseek-chat и deepseek-reasoner permanently retired |
One-liner: preview уже был сильным; GA добавляет targeted Agent/math/code upgrades плюс formal commercial billing. Underlying MoE architecture unchanged — GA про stability, optimization и pricing maturity.
Пропустить migration deadline: код с deepseek-chat сломается после 24 июля, 23:59 по Пекину.
Peak-hour blind calls: будни 09:00–12:00 и 14:00–18:00 по Пекину — rates ×2; batch jobs без off-peak scheduling быстро раздувают bill.
Нет Pro/Flash routing: V4-Pro на simple routing сжигает budget, который V4-Flash (output $0,28/M off-peak) закроет.
Игнор cache hits: без Prompt Cache на repeated system prompts теряете Flash cached input от $0,0028/M.
Benchmark misread: SWE-bench Verified 80,6% — open-source high, но Claude Fable 5 лидирует SWE-bench Pro (80,3%) — не выбирайте prod model по одному leaderboard.
V4-Pro vs V4-Flash: CSA, HCA, mHC и Muon — технический разбор
DeepSeek V4 даёт три architectural breakthroughs: KV-cache memory при 1M-token context падает до 10% от V3.2 (V4-Flash до 7%) — ultra-long context становится economically viable.
| Spec | V4-Pro | V4-Flash |
|---|---|---|
| Total parameters | 1,6 trillion (1,6T) | 284 billion (284B) |
| Active params per token | 49 billion (49B) | 13 billion (13B) |
| Transformer layers | 61 | 43 |
| Context window | 1 000 000 tokens | 1 000 000 tokens |
| Max output | 384K tokens | 384K tokens |
| Precision | FP4 (expert weights) + FP8 (rest) | FP4 + FP8 mixed |
| Pretraining data | 33T+ tokens | 32T+ tokens |
| License | MIT | MIT |
2.1 Hybrid attention: CSA + HCA
DeepSeek V4 заменяет V2/V3 MLA hybrid из двух новых attention mechanisms:
Compressed Sparse Attention: KV sequences Softmax-gated и pooled 4×; FP4 «lightning indexer» выбирает top-k sparse keys (Pro: top-1024, Flash: top-512) с 128-token sliding window. При 1M context inference FLOPs — 27% от V3.2.
Heavily Compressed Attention: tokens compressed 128×, затем global dense attention для long-range dependencies — complement к CSA. V4-Flash: first 2 layers HCA, далее CSA/HCA alternate; V4-Pro — similar pattern.
2.2 Manifold-constrained hyper-connections (mHC) и Muon optimizer
mHC апгрейдит standard residual connections: 4-channel residual stream + mixing matrices на Birkhoff polytope (doubly stochastic) — stable signals через 61 layers. Muon optimizer (вместо AdamW) применяет Newton-Schulz orthogonalization к gradients — faster convergence, stable training.
| Reasoning mode | Behavior | Best for |
|---|---|---|
| Non-think | No chain-of-thought; fastest response | Simple Q&A, routing |
| Think High | Explicit reasoning (thinking tags) | Medium complexity, code debugging |
| Think Max | Maximum reasoning; needs 384K+ context | Hard math, long-chain Agents |
Official sampling defaults: temperature=1.0, top_p=1.0 across all modes.
Benchmarks: SWE-bench, cost-performance и model comparison
| Benchmark | DeepSeek V4-Pro | Claude Fable 5 | GPT-5.6 Ultra | Claude Opus 4.8 |
|---|---|---|---|---|
| SWE-bench Verified | 80,6% (open-source #1) | 96,0% | Not separately published | ~69% |
| SWE-bench Pro | 55,4% | 80,3% | 78,1% | 69,2% |
| LiveCodeBench (Pass@1) | 93,5% | 88,1% | 87,4% | 83,2% |
| Codeforces Elo | 3 206 | — | — | — |
| Terminal-Bench 2.1 | 83,9% | 88,0% | 85,1% | 82,7% |
SWE-bench Verified — real GitHub bug fixes. 80,6% — current open-source high, tied с Gemini 3.1 Pro. SWE-bench Pro строже — Claude Fable 5 80,3% лидирует.
Artificial Analysis: Claude Fable 5 — $3,48 за Strategy & Ops index task (50 points); DeepSeek V4-Pro — $0,03 за comparable task (38 points) — 116× cheaper при ~12-point score gap.
| Dimension | DeepSeek V4-Pro | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|
| Open source / self-hostable | Да (MIT) | Нет (closed) | Нет (closed) |
| Context window | 1M tokens | Not publicly disclosed | 1M tokens |
| API output (off-peak) | $0,87/M | ~$15/M | $50/M |
| Peak output price | $1,74/M | — | — |
| Code capability | Very strong (near Fable 5) | Very strong | Strongest (SWE-bench Pro) |
| Data privacy | Да (private deploy) | Нет | Нет |
Scenario picks: tight budget / high volume / private deploy → V4-Pro или V4-Flash; max code quality, cost secondary → Claude Fable 5; hard algorithms + math → GPT-5.6 Sol / Ultra; massive log/document ingestion → V4-Flash (cached input from $0,0028/M).
Peak-valley pricing и migration steps до 24 июля
Главное GA billing change — peak-valley differentiated pricing, как time-of-use electricity. Peak hours (Beijing time): будни 09:00–12:00 и 14:00–18:00 — rates double.
| Model | Line item | Off-peak | Peak |
|---|---|---|---|
| V4-Pro | Input (cache hit) | ¥0,025 / $0,0035 / 1M | 2× |
| V4-Pro | Input (cache miss) | ¥3,00 / $0,435 / 1M | ¥6,00 / $0,87 |
| V4-Pro | Output | ¥6,00 / $0,87 / 1M | ¥12,00 / $1,74 |
| V4-Flash | Input (cache hit) | ¥0,02 / $0,0028 / 1M | 2× |
| V4-Flash | Input (cache miss) | ¥1,00 / $0,14 / 1M | ¥2,00 / $0,28 |
| V4-Flash | Output | ¥2,00 / $0,28 / 1M | ¥4,00 / $0,56 |
Four cost-saving tips: (1) non-real-time jobs off-peak (after 18:00 or before 09:00 Beijing); (2) Prompt Cache на repeated system prefixes; (3) lightweight → Flash, hard reasoning → Pro; (4) DeepSeek email alerts за 24h до billing changes. Даже в peak V4-Pro output $1,74/M — 8,6× cheaper vs Claude Opus 4.8 ($15/M).
Critical deadline: legacy deepseek-chat и deepseek-reasoner stop working 24 июля 2026, 15:59 UTC (24 июля, 23:59 Beijing).
| Legacy model | New model | Notes |
|---|---|---|
deepseek-chat | deepseek-v4-flash (non-think) | Fast; lightweight tasks |
deepseek-reasoner | deepseek-v4-flash (think mode) | Or upgrade to deepseek-v4-pro |
Six migration steps:
Search codebase: find every deepseek-chat и deepseek-reasoner reference.
Apply mapping: light chat → deepseek-v4-flash; former reasoner → Flash think mode or deepseek-v4-pro.
Update OpenAI SDK: change only model; keep base_url = https://api.deepseek.com.
Enable thinking (optional): extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}} replaces reasoner behavior.
Anthropic SDK compatibility: V4 supports OpenAI ChatCompletions + Anthropic Messages — update model only.
Staging validation: full regression before July 24; confirm zero legacy model names.
client.chat.completions.create(
model="deepseek-v4-pro",
messages=[...],
extra_body={"thinking": {"type": "enabled", "budget_tokens": 8000}}
)Стоит ли апгрейд на DeepSeek V4 GA? Citeable data и summary
DeepSeek V4 GA — один из key open-model milestones 2026. Value не в том, чтобы beat Claude Fable 5 или GPT-5.6 на каждом chart сегодня — а в undisputed open-source peak performance при ~1/10 to 1/100 closed flagship cost.
80,6% / open-source #1: SWE-bench Verified open-source high, tied с Gemini 3.1 Pro.
10% / 7%: at 1M tokens KV-cache memory — 10% of V3.2 (Flash as low as 7%).
$0,03 / 116×: V4-Pro Strategy & Ops task $0,03 vs Fable 5 $3,48 — 116× price gap, ~24% score difference.
Для teams без offshore data, indie devs на tight budget, Agent engineers с 1M context и AI products chasing cost efficiency — DeepSeek V4 Pro остаётся most balanced choice. Peak-valley billing adds planning overhead, но highly competitive — maturity signal, не retreat from value.
Альтернативы: DeepSeek Agent load tests на personal Mac — sleep и network interruptions на long-context jobs; closed APIs alone не match V4 MIT + $0,87/M off-peak output по cost и data sovereignty; miss July 24 migration cutoff — hard production outages. Для iOS CI/CD, DeepSeek V4 Agent 24/7 automation и local ds4 inference validation KVMNODE dedicated Mac Mini M4 cloud rental — usually better host: Apple Silicon unified memory, open sudo, flexible day/week/month terms. Подробнее: цены, центр помощи, оформить заказ.
Данные на 20 июля 2026 · Sources: DeepSeek official docs, arXiv:2606.19348, Hugging Face, Artificial Analysis