DeepSeek V4 Flash GA: 284B MoE бьёт 1.6T Pro на agent bench — post-training > scale

31 июля 2026 DeepSeek выкатила V4-Flash-0731 в official API beta — не новую архитектуру, а re-run post-training на том же 284B/13B active MoE stack. Vendor claim: TerminalBench 2.0 82.7 против 67.9 у V4-Pro preview, при $0.14 input cache-miss / $0.0028 cache-hit / $0.28 output per million tokens. Flagship V4-Pro GA и in-house agent framework Harness — still pending. Разбираем DSA sparse attention pipeline, Harness minimal-mode dependency, Artificial Analysis cost math vs Kimi K3 и Qwen3.8-Max, и 7-step pre-prod checklist для тех, кто реально деплоит agent loops.

SECTION 01 MoE math и пять red flags до первого prod deploy

После K3 dump 27 июля и Qwen3.8-Max GA 3 августа V4-Flash — третий китайский frontier за две недели. Для ML-инженера критично не перепутать marketing T-count с реальным compute footprint:

• Harness-dependent agent scores: TerminalBench 2.0 (82.7) и Toolathlon — measured через DeepSeek Harness minimal mode (not public yet): max reasoning, top_p 0.95, temperature 1.0. Changelog explicitly warns: agent benchmarks extremely sensitive to harness choice. Until third-party repro with Claude Code / Cursor — treat as vendor-plus-framework, not portable capability.
• V4-Pro GA + Harness ghost ship: Changelog says official Pro release and Harness «coming soon» — no confirmed date. Chinese media August 10–20 window = unnamed sources, not DeepSeek confirmed.
• API-only drop: 31.07 update hits API only — consumer app and web chat untouched. Don't extrapolate UX from API scores.
• Cache-hit reality: Overseas dev communities report low input cache-hit rates and occasional safety-classifier timeouts — compute budget still caps what post-training alone can fix.
• 「斩杀线」kill line: Chinese dev term for DeepSeek's combo of good-enough perf + rock-bottom price. Competitors must beat on capability OR undercut on price — explains GPT-5.6 Luna −80% same period.

SECTION 02 Release timeline и pricing matrix (snapshot 05.08.2026)

284B Flash beats 1.6T Pro on agent benches — signal that 2026 H2 post-training quality rivals raw parameter scale. Price vs Claude Opus 4.8: ~36× cache-miss input, ~179× cache-hit, ~89× output (media figures).

SECTION 03 Intelligence Index vs cost-per-task — third-party data

Artificial Analysis provides independent numbers separate from DeepSeek vendor agent runs. Core tension: V4-Flash isn't smartest in China, but cheapest per task.

Context: DeepSeek V4 GA deep dive, OpenRouter July rankings (Flash preview #1 call volume seven weeks pre-31.07).

SECTION 04 DSA stack, post-training pipeline и Harness framework

Technical report DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence describes three arch changes since April preview — independent of 31.07 post-training rerun:

• Hybrid DSA attention: CSA (Compressed Sparse Attention) + HCA (Heavily Compressed Attention) cut compute and KV-cache at 1M context. Vendor claim: V4-Pro needs 27% FLOPs and 10% KV-cache vs V3.2 — no independent repro yet. If accurate, million-token context becomes commercially viable at mainstream-length cost.
• mHC (Manifold-Constrained Hyper-Connections): improved residual flow for training stability — matters for fine-tuning teams on MIT weights.
• Muon optimizer: replaces classic optimizers for faster convergence — relevant for custom training loops on Apple Silicon adjacency (Metal shader compile chains still need native bare metal).

DeepSeek Harness — first named 31.07 — in-house agent execution framework: file I/O, tool calls, multi-step engineering. Positioned as Claude Code alternative. All published Flash-0731 agent scores run through Harness minimal mode. Production quality in Cursor/OpenCode may diverge significantly from changelog numbers.

SECTION 05 7-step pre-prod checklist + citable numbers + prod pattern

1. Audit legacy endpoints: grep codebase for deepseek-chat and deepseek-reasoner — retired 24.07. deepseek-v4-flash auto-points to 0731 build.
2. Monitor cache-hit rate: low hits jump input from $0.0028 to $0.14 per M — critical in full-history agent loops. Per-request logging mandatory.
3. Pre-plan peak-hour 2× surcharge: DeepSeek announced weekday Beijing 9–12 and 14–18 double rates — effective date unconfirmed. Batch scheduler ready.
4. Split benchmark sources: Artificial Analysis (Index 50, $0.03/task) separate from Harness TerminalBench (82.7) — different methodology.
5. Structured A/B vs K3 and Qwen: same prompts on code, long docs, agent tasks — K3 has local weights, Qwen API-only: document asymmetry.
6. Track Harness GA: tag internal agent scores «framework unknown» until public release — no exec SLA premature.
7. Split inference from compile layer: no LLM signs Xcode builds — pair with M4 cloud bare metal for iOS CI/CD.

• Arch: 284B total / 13B active MoE, 1M context, MIT license
• vs Claude Opus 4.8: ~36× cache-miss input, ~179× cache-hit, ~89× output
• Long-context efficiency: 27% FLOPs / 10% KV-cache vs V3.2 at 1M (vendor claim)
• OpenRouter: Flash preview #1 call volume seven weeks before 31.07
• Funding (background, unverified): media ~$7.4B round, ~$48.7B valuation — no official confirmation

Primary sources — re-check after upstream updates:

DeepSeek API — docs and changelog

Hugging Face — deepseek-ai/DeepSeek-V4-Flash

Artificial Analysis — Intelligence Index and task cost

V4-Flash-0731 — legit cost-efficiency play: 82.7 TerminalBench, $0.28/M output, MIT weights. But Harness-dependent scores, missing V4-Pro GA, API-only release = don't default prod without internal repro. Production pattern: DeepSeek API for cheap agent inference, VPSNIX M4/M4 Pro bare metal for native Apple Silicon builds — 100% Apple hardware, full root, zero hypervisor tax. Virtualized Mac instances suffer hypervisor loss, inconsistent Metal perf, fragile Xcode signing — unsuitable for 7×24 agent pipelines. Check pricing and keep inference and compile on separate stable layers.

SECTION 06 Часто задаваемые вопросы