2026 DEEPSEEK V4
ПОЛНЫЙ_
РЕЛИЗ_GA.
Тезис: Три месяца preview — и 20 июля 2026 DeepSeek V4 выходит в GA. Agent orchestration, math reasoning, code gen — затюнены под production. Новинка: пик/внепик тарификация (peak-valley pricing). Ниже — полный разбор: timeline, V4-Pro/Flash specs, CSA+HCA+mHC stack, три inference mode, benchmark tables, pricing grid, API migration (24 июля deadline), vs GPT-5.6 / Claude Fable 5, FAQ JSON-LD, deep insight, Mac/Metal angle.
30 секунд · TL;DR
| Релиз | 2026-07-20 GA · deepseek-chat EOL 24.07. |
| Scale | V4-Pro 1,6T (49B active) · V4-Flash 284B (13B active) · 1M context each |
| Benchmarks | SWE-bench Verified 80,6% (open-weight record) · LiveCodeBench 93,5% · MIT |
| Pricing | Peak-valley · V4-Pro output $0,87/M (off-peak) / $1,74/M (peak) |
| Cost ratio | 1/116 vs Claude Fable 5 ($0,03 vs $3,48/task) |
1. Pain points: зачем смотреть на полный релиз
- Preview был силён, production требовал SLA. С 24 апреля 2026 два вопроса: будет ли ещё performance? Изменится ли pricing? GA отвечает — три месяца production tuning + прозрачная peak-valley схема.
- Legacy API умирает через четыре дня.
deepseek-chatиdeepseek-reasonerотключаются 24.07.2026, 15:59 UTC. Cursor, OpenClaw, custom agents с старыми model strings — мигрируйте сейчас. - Peak-valley — инструмент load shaping, не просто price hike. Пекинские будни 09–12 и 14–18 — rates ×2. Но даже в peak V4-Pro output ($1,74/M) — 8,6× дешевле Opus 4.8 ($15/M). Ключ: off-peak scheduling + Flash routing.
2. Что это: DeepSeek V4 GA в одном абзаце
DeepSeek V4 GA — general availability от 20.07.2026: V4-Pro (1,6T MoE) и V4-Flash (284B MoE), MIT license, 1M token context window, 384K max output.
Vs preview: agent capabilities, math reasoning, code generation — production-tuned; peak-valley API pricing. Архитектура CSA + HCA + mHC — с апрельского preview; GA закрывает stability, tuning и commercial loop.
One-liner: лучший SWE-bench Verified среди open-weight (80,6%), near-frontier performance за 1/10–1/100 стоимости closed flagship, private deploy + реальный 1M context.
3. Timeline: preview → GA
| Дата | Событие |
|---|---|
| 2026-04-24 | V4 preview + open source (MIT), V4-Pro (1,6T) и V4-Flash (284B) |
| 2026-05 | Production-tuned builds, API GA |
| 2026-06 | V4-Pro permanent −75% (output $0,87/M) |
| 2026-06-29 | Email всем API users: GA mid-July, первое раскрытие peak-valley |
| 2026-07-19 | GA greyscale access, медиа «полный релиз завтра» |
| 2026-07-20 | GA live (дата этой статьи) |
| 2026-07-24 | deepseek-chat / deepseek-reasoner permanent EOL |
Контекст: DeepSeek custom chip & inference, обзор Kimi K3, AI скидки июнь 2026.
4. Model family: V4-Pro vs V4-Flash
| Spec | V4-Pro | V4-Flash |
|---|---|---|
| Total params | 1,6T | 284B |
| Active params/token | 49B | 13B |
| Transformer layers | 61 | 43 |
| Context window | 1 000 000 tokens | 1 000 000 tokens |
| Max output | 384K tokens | 384K tokens |
| Precision | FP4 (experts) + FP8 (rest) | FP4 + FP8 mixed |
| Pretrain data | 33T+ tokens | 32T+ tokens |
| License | MIT | MIT |
5. Архитектура: CSA + HCA + mHC — engineering deep dive
Классический Transformer: KV Cache растёт O(n) с context length — 1M tokens на inference silicon было экономически нереально. V4 выбрасывает MLA из V2/V3 и ставит три архитектурных хода:
5.1 CSA — Compressed Sparse Attention
- KV sequence сжимается 4× через Softmax-gated pooling
- FP4 «Lightning Indexer» для top-k sparse selection (V4-Pro: top-1024, V4-Flash: top-512)
- Sliding window последних 128 tokens
- На 1M context: 27% inference FLOPs vs DeepSeek V3.2 — это прямой выигрыш в tokens/sec на inference ASIC и Metal GPU alike
5.2 HCA — Heavily Compressed Attention
- Token compression 128×, global dense attention
- Long-range dependencies — complement к CSA
- V4-Flash: первые 2 слоя HCA, далее CSA/HCA alternate; V4-Pro — аналогично
Net effect: KV Cache memory = 10% от V3.2 на 1M (V4-Flash: 7%). Меньше HBM pressure → выше batch throughput на фиксированном silicon budget.
5.3 mHC + Muon optimizer
mHC (Manifold-Constrained Hyper-Connections) — upgrade residual path: x = x + f(x) заменяется 4-channel residual stream с mixing matrices на Birkhoff polytope (doubly-stochastic constraint). 61-layer deep network без gradient pathology.
Training: Muon optimizer вместо AdamW — Newton-Schulz orthogonalization градиентов, faster convergence, stable training at 1,6T scale.
5.4 Три inference mode
| Mode | Характеристика | Use case |
|---|---|---|
| Non-think | No CoT, min latency | Simple Q&A, routing |
| Think High | Explicit reasoning (<redacted_thinking>) | Mid complexity, code debug |
| Think Max | Max reasoning depth, 384K+ context | Hard math, long agent chains |
Sampling defaults: temperature=1.0, top_p=1.0 (official, all modes).
6. Benchmarks: open-weight leaderboard
| Benchmark | DeepSeek V4-Pro | Claude Fable 5 | GPT-5.6 Ultra | Claude Opus 4.8 |
|---|---|---|---|---|
| SWE-bench Verified | 80,6% ★ open-weight #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% |
Readout:
- SWE-bench Verified: V4-Pro 80,6% — open-weight top, tie с Gemini 3.1 Pro
- LiveCodeBench: V4-Pro 93,5% — ahead всех closed competitors
- SWE-bench Pro: Fable 5 80,3% — multi-file repo bugfixes
- Terminal-Bench: GPT-5.6 Ultra 85,1% — terminal/toolchain-heavy agents
Disclaimer: benchmarks — DeepSeek self-reported + third-party aggregation; independent reproduction in progress. Harness mismatch между моделями — учитывайте при сравнении.
7. Cost-per-task: economics at scale
Artificial Analysis, Strategy & Ops Index:
| Model | Cost/task | Score | Ratio |
|---|---|---|---|
| Claude Fable 5 | $3,48 | 50 pts | baseline |
| DeepSeek V4-Pro | $0,03 | 38 pts | 1/116 Fable 5 cost |
| DeepSeek V4-Flash | all 6 index tasks < $0,04 | — | light tasks optimal |
Fable 5 = 116× дороже за ~12 pts (31%) прироста. Для большинства production workloads V4-Pro — rational default.
8. Head-to-head: V4 vs GPT-5.6 Sol vs Claude Fable 5
| Dimension | DeepSeek V4-Pro | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|
| Open source | ✅ MIT | ❌ closed | ❌ closed |
| Self-host | ✅ | ❌ | ❌ |
| Context | 1M tokens | undisclosed | 1M tokens |
| Code | very strong (near Fable 5) | very strong | strongest tier |
| API output (off-peak) | $0,87/M | ~$15/M | $50/M |
| API output (peak) | $1,74/M | — | — |
| Agent capability | strong, multi-agent | strong | strongest tier |
| Data privacy | ✅ private deploy | ❌ | ❌ |
| Scenario | Pick | Why |
|---|---|---|
| Budget / high-frequency / self-host | V4-Pro / Flash | $0,87/M, MIT |
| Max code quality, cost irrelevant | Claude Fable 5 | SWE-bench Pro 80,3% |
| Algorithms + math reasoning | GPT-5.6 Sol / Ultra | Terminal-Bench lead |
| Massive log/doc processing | V4-Flash | cache hit input $0,0028/M |
9. Peak-valley pricing: full rate card
GA впервые вводит time-of-use differentiation — как electricity TOU rates. Peak (Пекин): будни 09:00–12:00 и 14:00–18:00, все rates ×2.
| Model | Line item | Off-peak | Peak |
|---|---|---|---|
| V4-Pro | Input (cache hit) | ¥0,025 / $0,0035 / 1M | ×2 |
| Input (cache miss) | ¥3,00 / $0,435 / 1M | ¥6,00 / $0,87 | |
| Output | ¥6,00 / $0,87 / 1M | ¥12,00 / $1,74 | |
| V4-Flash | Input (cache hit) | ¥0,02 / $0,0028 / 1M | ×2 |
| Input (cache miss) | ¥1,00 / $0,14 / 1M | ¥2,00 / $0,28 | |
| Output | ¥2,00 / $0,28 / 1M | ¥4,00 / $0,56 |
9.1 Четыре cost hacks
- Non-realtime → off-peak: doc batches, annotation, code review после 18:00 или до 09:00
- Prompt cache: repeated system prompts — V4-Flash cache hit от $0,0028/M
- Flash routing: intent на Flash, hard reasoning на Pro — 60–80% cost cut
- Billing alert: DeepSeek promises 24h email notice перед rate changes
Даже в peak: V4-Pro ($1,74/M) — 8,6× cheaper Opus 4.8 ($15/M) и GPT-5.6 Sol (~$15/M).
10. API migration (deadline 24 июля) ⚠️
Hard stop: deepseek-chat и deepseek-reasoner EOL 24.07.2026, 15:59 UTC (23:59 Пекин).
| Legacy | New | Note |
|---|---|---|
deepseek-chat | deepseek-v4-flash (non-think) | Fast, light tasks |
deepseek-reasoner | deepseek-v4-flash (think) | or deepseek-v4-pro for stronger reasoning |
10.1 OpenAI SDK (Python)
10.2 Anthropic SDK
V4 supports OpenAI ChatCompletions + Anthropic Messages — update model only:
11. Integration: пять путей в production
- Web/App: chat.deepseek.com — V4 GA default.
- Official API: platform.deepseek.com — OpenAI-compatible,
deepseek-v4-pro/deepseek-v4-flash. - OpenRouter: model ID
deepseek/deepseek-v4-pro— см. OpenRouter Mac routing guide. - Cursor / IDE: V4-Pro daily driver, fallback Claude Fable 5 для hard repo bugs.
- Codebase grep: search
deepseek-chat+deepseek-reasoner, staging test, prod cutover до 24 июля.
12. FAQ
Q: GA vs preview?
A: Agent/math/code improvements, peak-valley pricing, production stability. Architecture unchanged.
Q: V4-Pro vs V4-Flash?
A: Pro — complex reasoning, long agents; Flash — high-frequency routing ($0,28/M off-peak).
Q: Peak hours?
A: Пекин будни 09:00–12:00 и 14:00–18:00, rates ×2.
Q: deepseek-chat EOL?
A: 24.07.2026 15:59 UTC. Migrate to deepseek-v4-flash or deepseek-v4-pro.
Q: Self-host / local on Mac?
A: MIT, Hugging Face weights. V4-Pro needs GPU cluster; Mac for API or quantized V4-Flash via MLX.
Q: vs Claude Fable 5?
A: Fable 5 leads SWE-bench; V4-Pro at 1/116 cost with near-frontier perf + self-host.
13. Deep insight: open-source milestone signal 2026
DeepSeek V4 GA — не «preview graduation». Это structural inflection point в китайском open-source LLM ecosystem mid-2026.
13.1 От price disruptor к disciplined platform
18 месяцев DeepSeek держал рынок на ~$0,28/M output (V3 era), forcing OpenAI/Anthropic price cuts. «Почти бесплатно» unsustainable — compute, R&D, compliance требуют cash flow. Peak-valley = signal от subsidy → sustainable ops. Competitiveness не sacrificed: peak $1,74/M всё ещё 8,6× under Opus 4.8. Mechanism = demand-side scheduling, не cost passthrough. Teams с off-peak batch pipelines могут платить меньше, чем раньше.
13.2 Architecture redefines long-context economics
1M context в 2026 — commodity spec (Kimi K3, Fable 5). V4 value prop: 1M context economically servable. CSA 4× + top-1024 + 128-token window → 27% FLOPs, 10% KV Cache vs V3.2. Kimi K3 $15/M output vs V4-Pro $0,87/M — order-of-magnitude gap на million-token workloads.
mHC 4-channel residual + Muon = stable training 61 layers at 1,6T. Parameter jump 671B → 1,6T без training collapse.
13.3 MIT license — strategic weight
July 2026: license = parameter count в importance. Llama commercial restrictions, «research only» clauses — compliance landmines. MIT на V4:
- Unlimited on-prem, no royalty, no reporting
- Finetune, distill, ship closed-source products
- Data residency для gov/finance/health без US API egress
Kimi K3 claims 2,8T «largest open» (см. K3 обзор); V4 competes on engineering efficiency, cost structure, license freedom. Для Mac dev: MIT + dual SDK (OpenAI/Anthropic) = seamless switch в Cursor, Continue, OpenClaw; MLX/llama.cpp quantization path open. V4-Flash 13B active на Apple Silicon unified memory — community quant experiments running; track Apple Silicon V4 AMX-2 benchmarks для local tokens/sec и Metal throughput numbers.
Bottom line: V4 GA не бьёт Fable 5 / GPT-5.6 на every leaderboard — yet. Но delivers undisputed strongest open-weight performance at 1/10–1/100 closed flagship cost. Для teams без data egress, budget-constrained devs, 1M-context agent engineers — V4-Pro = no-compromise pick.
14. Closing: API anywhere — agent validation on Mac/Metal
platform.deepseek.com API key, OpenRouter route swap — Windows/Linux достаточно. Но если вам нужен Cursor + V4-Pro long-context coding, Think Max stress на multi-repo SWE-bench regression, 1M doc batch jobs, или MLX quantized V4-Flash validation перед prod deploy — Apple Silicon unified memory + Metal compute stack = lowest-friction path. Zero PCIe bottleneck на KV-heavy workloads vs discrete GPU + system RAM split.
Pragmatic split: daily driver на API, V4-Pro Think Max load tests, SWE-bench regression suites, 1M context doc batches — на MACGPU remote Mac mini M4 node. On-demand, SSH-secured, peak-valley cron scheduling на remote silicon. Main machine остаётся stable для everyday dev; Metal throughput на remote node съедает agent memory pressure.