Qwen3.8-Max GA: 2.4T sparse MoE, Arena Text #5 — weights promise broken (пока)

3 августа 2026 Alibaba выкатила Qwen3.8-Max в GA — крупнейший Qwen на сегодня. Под капотом: 2.4T total params, 95B active per forward pass (sparse MoE + hybrid attention на базе Qwen 3.5), 1M token context (input cap 1 048 576, output ~131K), multimodal backbone. API pricing: $2 input / $6 output per million tokens. Arena Text — #5 с 1496 preliminary score, Vision — #2. Но weights, которые обещали «на следующей неделе», на 4 августа всё ещё не на Hugging Face / ModelScope — и почти все топовые цифры (TerminalBench 86.6, PaperBench 93.0, OSWorld-Verified 86.1) — vendor-run, без independent audit. Разбираем arch stack, сравнение с Kimi K3 и DeepSeek V4, transparency drama и 7-step integration checklist для тех, кто реально деплоит.

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

После K3 full-weight dump 27 июля и V4 GA 20 июля, Qwen3.8-Max — третий китайский frontier за две недели. Для ML-инженера критично не перепутать marketing T-count с реальным compute footprint:

• 2.4T ≠ inference cost 2.4T: active experts — 95B. Сравнивать с K3 (2.8T total, 16 active experts) или V4-Pro dense 1.6T без колонки «active» — классическая ошибка в TCO-моделях.
• Weights ghost ship: Alibaba анонсировала first-ever Max-class open weights на HF/ModelScope «next week» — 4.08.2026 upload отсутствует. Self-host pipeline, quantization plan и license review — blocked.
• Dual benchmark regime: Arena (#5 text, #2 vision, 1496 preliminary) — third-party crowd eval. TerminalBench/PaperBench/OSWorld — internal harness, reproducibility unknown. Не смешивать в одном dashboard.
• 1M context billing trap: теоретический 1M window при $2/M input звучит дёшево, но agent loop с full-history replay сжигает budget. Cache policy Model Studio vs OpenRouter routing — must benchmark.
• Product surface fragmentation: Model Studio API ≠ QwenWork office agent ≠ Apple Intelligence China Qwen backend (CAC registered 15.07). Три разных data residency и ToS.

SECTION 02 GA timeline и spec sheet (snapshot 03.08.2026)

1496 Arena points + $2/$6 API — solid mid-tier frontier positioning. Без weights и без independent code eval это всё ещё «trust me bro» tier, не production default.

SECTION 03 Head-to-head: Qwen3.8-Max vs Kimi K3 vs DeepSeek V4-Pro

Все три заявляют 1M context. Различия — в weight availability, $/token и benchmark methodology:

Deep dive: Kimi K3 review, OpenRouter July rankings, K3 weights release.

SECTION 04 QwenWork agent layer, Apple Intelligence CN и transparency controversies

GA day Alibaba включила QwenWork (千问办公) — office agent platform поверх Qwen3.8-Max: slides, spreadsheets, long research, 16-day autonomous SWE demo (official), 500+ step chip design optimization showcase. Это SaaS product, не raw API endpoint — отдельный DPA, log retention, region pinning.

Parallel track: CAC registration 15.07.2026 — China-market Apple Intelligence uses Qwen for generation, Baidu for search/Siri. Qwen3.8-Max GA — logical capability bump candidate для iOS 27, но ship timeline на стороне Apple integration team. Cross-border apps: test matrix global backend vs CN backend обязателен.

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

1. Block calendar for weight drop: until HF/ModelScope live — tag internally as API-only, no self-host commits to stakeholders.
2. Split benchmark sources: Arena rows separate from Alibaba TerminalBench/PaperBench — log measurement date + harness version.
3. Run 1M context cost sim: representative agent loops at $2/$6; compare Model Studio vs OpenRouter fallback routing.
4. Structured A/B vs K3 and V4: same prompt set — frontend code, long doc summary, vision input. K3 has local weights, Qwen API-only: document asymmetry.
5. Define QwenWork vs custom agent boundary: data residency, Alibaba Cloud region, log retention — enterprise DPA required for SaaS path.
6. Log transparency risks in selection memo: unfulfilled weight promise, vendor-run benches, preliminary Arena — prevents exec overcommit.
7. Split inference from compile layer: no LLM signs Xcode builds — pair with M4 cloud bare metal for iOS CI/CD.

• Params: 2.4T total / 95B active MoE
• Context: 1M in · ~131K out
• Arena: Text #5 (1496 prelim) · Vision #2
• API: $2 / $6 per M tokens
• Weights: not shipped 08-04; HF/ModelScope promised
• Competitor API: K3 $3/$15 · V4-Pro off-peak out $0.87/M
• Products: QwenWork office agent · Apple Intelligence CN (Qwen gen backend)

Primary sources — re-check after upstream updates:

Alibaba Cloud — Qwen3.8-Max GA press release

Alibaba Cloud Model Studio — API docs

Arena.ai — Text / Vision leaderboards

Qwen3.8-Max — legit mid-tier frontier play: Arena #5, Vision #2, sane $2/$6 API. Но missing weights + vendor-run benchmarks + broken「next week」promise = don't default prod without internal repro. Production pattern: Model Studio или future private endpoint для long-context inference и QwenWork-style agents, VPSNIX M4/M4 Pro bare metal для native Apple Silicon builds — 100% Apple hardware, full root, zero hypervisor tax, elastic daily/monthly billing. Check pricing и держи long-running agents на compliant physical infra.

SECTION 06 FAQ