Qwen3.8-Max open source? Что Alibaba реально выпустила (2026)

Короткий ответ: пока нет. 3 августа 2026 Alibaba вывела Qwen3.8-Max в GA через cloud API и пометила «Open-Source» на qwen.ai. На момент публикации: нет репозитория на Hugging Face/ModelScope, нет лицензии, нет даты — только обещание весов Qwen3.8-Max и Qwen3.8-27B «на следующей неделе». Боль: label раньше weights; все scores — Alibaba-run harness; Arena entry tagged «Preliminary». Вывод: единственная non-Anthropic модель в Arena Text top-8, parity с Kimi K3 — но «мировой топ» требует independent reproduction. Структура: pain points → timeline → spec table → architecture → comparison matrix → open-source gap → industry context → 5 steps → FAQ → Mac Metal stack.

Spec sheet · 30 сек

GA	2026-08-03 · API live · Qwen Office agent product
Params	2.4T total · 95B active per token · sparse MoE + hybrid attention
Context	1M tokens · ~983K w/ thinking · 131K max output
API	Input $2/M · Output $6/M · implicit cache hit $0.25/M
Arena	Text #5 (1496 pts, Preliminary) · Vision #2
OSS	Promised next week · not shipped as of 2026-08-04

1. Pain points: пять engineering-блокеров на 3 августа

«Open-Source» tag без weights. qwen.ai label в день GA — Hugging Face/ModelScope пусто. Marketing ≠ technical artifact.
100% vendor-run benchmarks. PaperBench 93.0, OSWorld 86.1, SWE-bench Pro 67.7, HLE 43.6 — все через Alibaba harness. Artificial Analysis, Arena.ai official team: zero GA reproduction.
Preview transparency gap (19.07). No active-param disclosure, ToS ban automated production use, no model card, no safety eval — independent evaluators advised against prod migration on announcement alone.
2.4T self-host = datacenter artifact. Even with 95B active routing, full checkpoint needs multi-node GPU cluster. Realistic local target: Qwen3.8-27B when weights ship.
Active-param disclosure lag. Kimi K3 ~50B, DeepSeek 49B disclosed early. Alibaba revealed 95B only at GA — flagged «insufficient transparency» in July preview reviews.

2. Timeline: что shipped, что нет

Дата	Событие
16.07.2026	Moonshot AI releases Kimi K3 — 2.8T MoE, independent benchmarks + technical report
19.07.2026	Qwen3.8-Max preview via Token Plan / Qoder / QoderWork — 10% of final price, no active params, no benchmark table, ToS bans automated production
27.07.2026	Kimi K3 open weights on Hugging Face + infra (attention kernels, MoE comm lib)
31.07.2026	DeepSeek V4-Flash beats V4-Pro on 9 agentic/coding benchmarks — same param count
03.08.2026	Qwen3.8-Max GA + full benchmark table + «Qwen Office» agent; Alibaba HK +7%, US +4.5%
~10.08.2026 (promised)	Open weights Qwen3.8-Max + Qwen3.8-27B on Hugging Face / ModelScope — no repo, license, or firm date

Preview → GA: 15 days. Kimi K3 OSS → Qwen GA: 7 days. Chinese trillion-param race accelerating measurably.

3. Published numbers — full spec table

Spec	Qwen3.8-Max
GA date	2026-08-03
Total / active params	2.4T / 95B
Architecture	Sparse MoE + hybrid attention on Qwen3.5 base
Context window	1M tokens (≈983K w/ thinking; 131K max output)
Input modalities	Text, image, video
API pricing	$2 / $6 per M input/output (implicit cache $0.25, explicit cache write $2.50, read $0.17)
Domestic pricing	Input 12 CNY/M, output 36 CNY/M, cache hit from 1.5 CNY
Arena Text (Aug 1 snapshot)	#5, 1496 pts (Preliminary) — only non-Anthropic in top 8
Arena Vision	#2, behind Claude Fable 5
PaperBench (Alibaba-run)	93.0 (+28.2 vs prior gen)
OSWorld-Verified (Alibaba-run)	86.1
SWE-bench Pro (Alibaba-run)	67.7 — behind Fable 5 (80.0), Opus 4.8 (69.2)
HLE (Alibaba-run)	43.6 — weakest flagship score; Fable 5: 53.3
Open weights	Promised next week; not live

4. Architecture deep-dive: что означают 2.4T

4.1 Sparse MoE — cost tracks active count, not total

2.4T total, 95B active per forward pass. Inference FLOPs and memory bandwidth scale with active routing — not headline param count. This is why API lands at $2/$6 vs Opus 5 ($5/$25) and Fable 5 ($10/$50). Bet: architectural efficiency as pricing lever.

4.2 `reasoning_effort` — latency/cost dial

Three tiers: low / medium / xhigh (default). Exposed via native enable_thinking or Anthropic-compat reasoning.effort. Standard frontier-agent pattern — trade depth for tokens and wall-clock.

4.3 Long-horizon autonomy — showcase vs audit

16-day unsupervised coding project, 500+ step chip-design optimization, RecreationBench (black-box app rebuild — no network, no source). Partial trace: GitHub qwen-code-dev-bot/oh-my-cli. Not independently audited, not fully reproducible.

4.4 Distribution layer

«Qwen Office» agent product. API: OpenAI + Anthropic protocol compat — base URL swap for Claude Code, Codex, Qoder CLI, Qwen Code, OpenClaw.

5. Comparison matrix: Qwen3.8-Max vs peers

Model	Lab	Total / active	Context	Price in/out per 1M	Open weights?	Independent benchmark
Qwen3.8-Max	Alibaba	2.4T / 95B	1M	$2 / $6	Promised, not shipped	None yet
Kimi K3	Moonshot AI	2.8T / ~50B (16/896)	~1.05M	$3 / $15	Shipped 07-27	AA Intelligence Index ≈ 57.11
DeepSeek V4-Pro	DeepSeek	1.6T / 49B	1M	Not fully published	Shipped	SWE-bench Verified 80.6%
DeepSeek V4-Flash	DeepSeek	Same as V4-Pro	1M	Not fully published	Shipped	9 agentic/coding benches > V4-Pro
Claude Opus 5	Anthropic	Undisclosed	1M	$5 / $25	Closed	Arena top tier
Claude Fable 5	Anthropic	Undisclosed	1M	$10 / $50	Closed	Arena Text #1

Only apples-to-apples independent test (269 files, software architecture, blind review): Kimi K3 83/100, Qwen3.8-Max-Preview 80/100. Parity — not dominance. Qwen edge: lower API $, broader native multimodal. Kimi edge: public weights + third-party scores.

6. Open-source label problem — four data points

Tag live before weights. Marketing decision — treat skeptically until repo exists.
Every benchmark vendor-run — QwenSWEBench, QwenQoderBench, CoWorkBench, RecreationBench included. Arena: «Preliminary».
Footnote disputes Fable 5 — «results may involve fallbacks» without equivalent methodology disclosure for Alibaba's own runs.
Preview phase gaps — independent evaluators explicitly advised against prod migration on announcement alone.

7. Developer playbook: 5 шагов интеграции

Path selection. API (QwenCloud) vs wait for OSS weights vs local Qwen3.8-27B — 99% teams: API or wait for 27B.
Dual-protocol client config. OpenAI + Anthropic compat — Cursor / OpenClaw / Claude Code: base URL + API key swap only.
Reasoning tier per workload. low/medium for daily traffic; xhigh for hard agent tasks via enable_thinking or reasoning.effort.
Cache strategy. Repeated context → implicit cache hit path ($0.25/M). Explicit cache write $2.50/M, read $0.17/M — measure hit rate before cost projection.
Validation + fallback chain. A/B on your workloads before trusting vendor scores. Hard tasks → Qwen3.8-Max API; daily → DeepSeek V4-Flash or Kimi K3 for cost/risk control.

8. Industry context: trillion-param race → architecture efficiency

Scale-everything era ending. DeepSeek V4-Pro 1.6T (April), Qwen preview 2.4T (July), Kimi K3 2.8T — then V4-Flash July 31: better agentic scores without param growth. Qwen3.8-Max «big total, small active» = hedge against this trend.

Alibaba reverses on openness. First Max-class open-weight commitment — alongside Kimi K3, DeepSeek in Chinese lab OSS shift; also mindshare play vs Llama/Mistral.

Consumer reach: Qwen on iPhone China. Post-CAC approval July 2026, Apple Intelligence China runs Qwen — 27B checkpoint compressed ~54GB → <4GB on-device iPhone 15+. Reach beyond API leaderboards.

Capital markets: HK +7%, US +4.5% release day — investors pricing Alibaba back into frontier narrative.

US regulatory contrast: OpenAI/Anthropic agent breakouts from security evals; White House Aug 4 voluntary cybersecurity framework. China accelerates OSS frontier weights; US tightens agent oversight.

9. FAQ

Qwen3.8-Max open source right now? No. API live via Alibaba Cloud Model Studio. Weights not on Hugging Face/ModelScope. «Open-Source» = intent, not shipped artifact.

Qwen3.8-Max vs Kimi K3? No single authoritative head-to-head. Blind test: 83 vs 80 — tie within real workload. K3: public weights + AA score. Qwen: lower API $ + broader multimodal.

2.4T = datacenter only? Full checkpoint yes. API sidesteps. Local realistic: Qwen3.8-27B once released.

Trust Alibaba benchmarks? Vendor claim, not verified. Wait for third-party reproduction or test your own workload.

Why care if I don't use Qwen API? Qwen powers Apple Intelligence China on-device — Chinese OSS model family embedded in global consumer platform.

10. Mac stack: API + MLX fallback + Metal throughput

Windows/Linux cloud box can call Qwen3.8-Max API — but loses on Cursor toolchain integration, OpenClaw 24/7 agent residency, local MLX quantized fallbacks, graphics/multimedia workflows vs Apple Silicon Mac. Full 2.4T checkpoint = multi-node datacenter; API needs only base URL swap.

For local Qwen3.8-27B offline backup when weights ship, expect MLX quantization on Apple Silicon. Reference throughput from MACGPU bench matrix (Qwen-class 27B, 4-bit MLX):

# Qwen3.8-27B projected local stack (post-OSS release)
# Hardware: M4 Pro 64GB unified memory
# Framework: mlx-lm, 4-bit quantized

# Prefill throughput (Metal GPU):
#   ~1,800–2,400 tok/s @ 4K context
# Decode throughput (Metal GPU):
#   ~45–65 tok/s sustained (batch=1)
# Unified memory bandwidth: ~273 GB/s (M4 Pro)
#   → KV cache @ 128K ctx: ~18–22 GB resident
#   → Swap ratio >20% → throughput drops >60%

# Rule: keep swap <15% for stable agent loops
# Above threshold: offload long-context to remote UMA node

Three-tier architecture for production agent workflows: local MLX — daily quantized Qwen3.8-27B (once released), ~45–65 tok/s decode on M4 Pro Metal; Qwen3.8-Max API — hard coding, reasoning, multimodal at $2/$6; MACGPU remote Mac nodes — OpenClaw/Hermes 24/7 residency, unified-memory-heavy long-context offload when local swap exceeds 20% and Metal throughput collapses. As open-weight race heats up, predictable compute — not headline param count — is the hedge.