2026 DeepSeek V4 Flash официально: бенчмарки, цена и сравнение с Claude Opus 4.8
31 июля 2026 DeepSeek перевёл V4-Flash-0731 в официальный API-статус без изменения весов: 284B total / 13B active, контекст 1M tokens, MIT. Agent-benchmarks по vendor data обгоняют V4-Pro preview в 5,7× по параметрам; list price $0,14 / $0,28 за M tokens — до 179× ниже Claude Opus 4.8 на cache hit. Разбор CSA+HCA, KV-cache footprint, Harness caveats и матрица API vs ds4/Metal на Apple Silicon UMA.
1. Timeline: preview апреля → официальный build 0731
- 24 апреля 2026: V4 preview — V4-Pro (1,6T/49B active), V4-Flash (284B/13B), 1M context, MIT, Hugging Face weights.
- 24 июля 2026:
deepseek-chatиdeepseek-reasonerretired; весь traffic на V4 namespace. - 27 июля 2026: Moonshot AI — Kimi K3 (2,8T) open weights, ~104B active (community estimate).
- 31 июля 2026:
deepseek-v4-flashbuild 0731 official API beta; weights sync HF. Changelog: первое упоминание DeepSeek Harness. API-only — app/web без изменений. - На 05.08.2026: V4-Pro GA — «as soon as possible», даты нет.
2. Три production bottleneck для inference-инженеров
- Harness-dependent scores ≠ portable capability: Terminal Bench 2.0 — 82,7 vs V4-Pro preview 67,9 — измерено Harness minimal mode (max, top_p=0,95, temp=1,0), framework не публичен. Портировать на Claude Code/Cursor без independent reproduction — ошибка capacity planning.
- KV-cache и memory bandwidth ceiling: при 1M context vendor claim: V4-Pro = 27% FLOPs и 10% KV footprint vs V3.2. На Mac с UMA это критично: 128 GB host с 81 GB q2 weights оставляет ~100–300K usable context; 512 GB Ultra — единственный realistic tier для full 1M production. Low cache-hit rate в API усиливает cost multiplier.
- Peak-hour pricing и classifier latency: DeepSeek анонсировал 2× surcharge 09–12 / 14–18 Peking; overseas devs report safety-classifier timeouts. Batch без temporal scheduling и fallback chain — SLA killer.
3. Ценовая и модельная матрица
Vendor list price за M tokens (05.08.2026), не independent audit.
| Модель | Статус | Total / Active | Input (miss/hit) | Output | License |
|---|---|---|---|---|---|
| DeepSeek V4-Flash-0731 | Official (31.07) | 284B / 13B | $0,14 / $0,0028 | $0,28 | MIT |
| DeepSeek V4-Pro | Preview only | 1,6T / 49B | $0,435 / $0,003625 | $0,87 | MIT |
| Kimi K3 | Open weights (27.07) | 2,8T / ~104B | $3,00 / $0,30 | $15,00 | Modified MIT |
| GLM-5.2 | Open (June 2026) | ~744B / ~40B | n/a | n/a | MIT |
| Qwen3.8-Max | API GA (02.08) | 2,4T / 95B | $2,00 / ~$0,17–0,25 | $6,00 | Open promised |
| Claude Fable 5 | Closed | n/a | n/a | n/a | Proprietary |
4. Benchmarks: vendor vs Artificial Analysis
| Модель | Intelligence Index (AA) | Avg cost/task (AA) | Terminal Bench 2.0 (vendor) |
|---|---|---|---|
| V4-Flash-0731 | 50 | $0,03 | 82,7 (Harness minimal) |
| Kimi K3 | 57 | $0,86 | n/a |
| GLM-5.2 | ~51 | n/a | n/a |
| GPT-5.6 Sol | 9+ pts above | $1,86 | n/a |
| Claude Fable 5 | 9+ pts above | $3,15 | n/a |
V4-Flash не index leader — Kimi K3 и GLM-5.2 выше. DeepSeek играет в «good-enough intelligence × rock-bottom price»: cost/task ~1/29 Kimi, ~1/62 GPT-5.6 Sol, ~1/105 Claude Fable 5. Preview семь недель #1 на OpenRouter по call volume.
5. Архитектура: DSA, mHC, Muon и throughput при 1M context
Ключевой факт: V4-Flash-0731 byte-identical architecture к april preview. Performance jump — чистый post-training; 284B/13B MoE beats 1,6T/49B sibling на agent tasks.
Technical report «DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence»:
- Hybrid attention (DSA): Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) — снижение compute и KV RAM на long context.
- mHC: Manifold-Constrained Hyper-Connections — улучшенные residual paths, стабильнее gradient flow на deep stacks.
- Muon optimizer: faster convergence vs Adam-class на large MoE training runs.
Vendor metrics @ 1M tokens (V4-Pro vs V3.2): 27% per-token FLOPs, 10% KV-cache footprint. Independent reproduction пока не опубликована — но если верно, million-token context становится commercially viable, не marketing spec.
Локальный inference на Metal (ds4): antirez ds4 — pure C, Metal backend first-class. Flash q2 ~81 GB weights; M3 Ultra 512 GB: README benchmark 468 t/s prefill / 27 t/s generation на 11,7K prompt. UMA eliminiert PCIe hop при MoE expert routing — zero-copy expert selection. Disk-resident KV (--kv-disk-dir) критичен для session resume без full prefill.
DeepSeek Harness: in-house agent runtime (file I/O, tool calls, multi-step engineering) — DeepSeek answer to Claude Code. Все published agent scores — Harness minimal mode, not yet public release.
6. Пять шагов: API, ds4/Metal и agent routing
- Выбрать deployment tier. Cloud API
deepseek-v4-flashдля batch agents без data residency constraints. Self-host: ds4 + Metal на 128 GB (Flash q2) или 512 GB Ultra (q4 / parallel sessions). Complex reasoning — V4-Pro preview до GA. - Legacy migration. Grep codebase/CI на
deepseek-chat/deepseek-reasoner; switch todeepseek-v4-flash— aliases dead since 24.07.2026. - OpenClaw tiered routing. Tier 0 batch → V4 Flash; Tier 1 coding → Kimi K3; Tier 2 premium → Claude Opus 5. Primary + 2 fallbacks в
openclaw.json, 429 auto-switch. - Peak-hour + cache. Heavy batch outside 09–12 / 14–18 Peking; Prompt Cache для system prompts ($0,0028/M hit). Локально: schedule prefill-heavy jobs когда thermal headroom max (ночной cron на Remote Mac).
- 24/7 gateway + ds4 co-location.
openclaw gateway install+ launchd на Remote Mac; ds4-server с disk KV на NVMe; workspace sync SFTP/rsync.
# ds4-server на Apple Silicon (Metal)
git clone https://github.com/antirez/ds4 && make metal
./ds4-server --ctx 100000 \
--kv-disk-dir /Volumes/Data/ds4-kv \
--kv-disk-space-mb 8192
# OpenAI-compatible API call
curl https://api.deepseek.com/v1/chat/completions \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"..."}]}'7. FAQ
Q: V4 Flash или V4 Pro? Batch agents, cron, cost-sensitive pipelines → V4-Flash-0731 (beats V4-Pro preview на vendor agent benches). Deep world knowledge / complex reasoning → V4-Pro preview until GA.
Q: Open source? Да — MIT на Hugging Face (deepseek-ai/DeepSeek-V4-Flash), commercial use без доп. license.
Q: V4-Pro GA когда? Нет confirmed date. Changelog: «as soon as possible». August 10–20 — unverified rumors.
Q: Harness доступен? Нет — только changelog mention; scores from internal minimal mode.
Q: 1M context на Mac realistic? Full 1M KV ~26 GB по README; 128 GB host caps ~100–300K practical; 512 GB Ultra для production long-context.
8. Итог: post-training побеждает scale — но bottleneck остаётся hardware
V4-Flash-0731 доказывает: H2 2026 competition смещается от parameter arms race к post-training quality. «Good enough + extreme price» задаёт kill line для commodity agent workloads — preview seven weeks #1 OpenRouter не случайность.
API path решает cost, но не throughput ceiling: MoE routing на fragmented GPU memory убивает long-context prefill; laptop thermal throttle маскируется под «model regression». ds4 + UMA + disk KV на dedicated Apple Silicon node — единственный path к reproducible t/s без cloud egress.
Laptop sleep, overloaded VPS и missing fallback chains уничтожают ROI быстрее любого price cut. SFTPMAC Remote Mac — 128/256/512 GB Apple Silicon tiers под ds4 и OpenClaw gateway: launchd uptime, NVMe под disk KV, SFTP/rsync workspace, OpenRouter low-latency routing. Pay-per-hour вместо $13K Studio capex; switch memory tier с model roadmap. Для команд, которые serious про agent inference на Metal — стабильнее home PC и дешевле 24/7 self-hosted rack.