2026年8月15日星期六

Qwen3.8-Max Review 2026: Alibaba's 2.4T Open-Source Frontier, Benchmarks & Pricing

Qwen3.8-Max Review 2026: Alibaba's 2.4T Open-Source Frontier, Benchmarks & Pricing

Qwen3.8-Max Review 2026: Alibaba's 2.4T Open-Source Frontier, Benchmarks & Pricing

Verdict first: Qwen3.8-Max is the most consequential open-weight release of 2026 — a 2.4 trillion-parameter Mixture-of-Experts giant that matches closed flagships on paper for roughly one-fifth of the price. The open weights (the first Max-class model Alibaba has ever open-sourced) make it a genuine frontier contender for research, enterprise and price-sensitive agent workloads. Just treat the self-reported benchmarks with caution: independent testing shows the top score comes with 64 turns and 40% hallucination.

Qwen3.8-Max review 2026 — Alibaba's 2.4T-parameter open-source MoE flagship

Qwen3.8-Max — Alibaba's largest and most capable model, now with open weights

What Is Qwen3.8-Max?

Qwen3.8-Max is Alibaba Cloud's flagship large language model, unveiled on August 3, 2026. It is the largest and most capable model in the Qwen family — a sparse Mixture-of-Experts (MoE) network with 2.4 trillion total parameters and ~95 billion activated per token (about 4% activation).

It went live on Alibaba Cloud Model Studio the same week, and on August 12, 2026 Alibaba published the open weights (Qwen3.8-2.4T-A95B) on Hugging Face and ModelScope — the first time it has ever open-sourced a Max-class flagship. The announcement alone lifted Alibaba's stock by roughly 130 billion yuan in market value.

Why it matters: Qwen3.8-Max is the strongest sign yet that China's open-weight ecosystem can sit at the frontier. It ranks #2 globally in vision (Arena AI, 1305) and #5 in text (1496), trailing only Claude Fable 5 among the models tested — while being dramatically cheaper than US closed flagships.

What makes it different from Qwen 3.7?

This is a generation jump, not a point release. Compared with Qwen 3.7, Alibaba reports a +28.2 point jump on PaperBench (93.0 vs ~65) and a +32.8 point jump on FrontierSWE (73.5 vs 40.7). It also carries a much larger context (roughly 1M tokens vs 256K) and a far bigger parameter count (2.4T vs ~1T).

Key Specs at a Glance

SpecQwen3.8-Max
ReleasedAugust 3, 2026 (API) · August 12, 2026 (open weights)
ArchitectureSparse MoE — 2.4T total params, ~95B active per token (~4%)
Context window~1M tokens (991K no-thinking / 983K thinking mode)
Max output131,072 tokens
Thinking budgetUp to 262K tokens
Input modalitiesText, image, video (API) · text-only (open weights)
Open weightsQwen3.8-2.4T-A95B on Hugging Face / ModelScope
Pricing$2.00 / $6.00 per M tokens (in / out), $0.25/M cache
Sister modelQwen3.8-27B (smaller, local-friendly)

Benchmarks: Real or Hype?

Qwen3.8-Max posts frontier-level numbers, but every headline figure is self-reported by Alibaba — none had independent verification at release. Here's the picture:

BenchmarkQwen3.8-MaxContext / Rival
PaperBench93.0Strongest result; ahead of GPT-5.6 Sol & Fable 5 (vendor-reported)
GPQA Diamond92.6Graduate-level science reasoning
OSWorld-Verified86.1Computer-use / agent tasks
Terminal-Bench 2.186.6GPT-5.6 Sol leads at 88.8
FrontierSWE73.5Up from 40.7 in prior generation
SWE-bench Pro67.7Trails Fable 5 (80.0) & Opus 4.8 (69.2)
Humanity's Last Exam43.6Lowest among flagship models — a clear weak spot

Arena AI (independent leaderboard): #2 in vision (1305), #5 in text (1496) — a genuinely frontier placement.

The catch (independent testing by Artificial Analysis): Qwen3.8-Max reaches an Agentic Index of 58, on par with Claude Opus 5 — but it takes 64 turns per task (vs 14 for its predecessor) and ~$1.14 per task, and its measured hallucination rate rises from 23% to 40%. The score is real; the efficiency and reliability behind it are not.

Analysts also flagged that the multimodal comparison table was measured against the weaker Qwen3.7-Plus rather than the stronger Qwen3.7-Max, flattering the improvement. Treat "93.0 PaperBench" as an upper-bound claim until third-party runs reproduce it.

Pricing & Value

ModelInput / M tokensOutput / M tokensCache
Qwen3.8-Max$2.00$6.00$0.25
GPT-5.6 (OpenAI)$2.00$6.00
Claude Opus 5$5.00$25.00
Claude Fable 5$10.00$50.00
Kimi K3 (Moonshot)$3.00$15.00

Value verdict: $2/$6 per M tokens is a frontier price, not a premium price — identical to GPT-5.6 and about 5× cheaper than Claude Fable 5 on input, and 33% cheaper on input than Kimi K3. At this price point, Qwen3.8-Max is the cheapest way to run near-frontier capability on a per-token basis (even more so if you self-host the open weights).

Open Source & Self-Hosting

The headline story: Qwen3.8-Max is now open. On August 12, 2026, Alibaba dropped the Qwen3.8-2.4T-A95B open weights on Hugging Face and ModelScope — the first Max-class flagship the company has ever open-sourced, just days after a wave of domestic models claimed to match Anthropic's performance.

  • What you get: full weights of the 2.4T MoE flagship (95B active), text-only, under an open license (same family as prior Apache-style Qwen releases — verify the exact license on the model card before commercial use).
  • What you don't get: the open version is text-only (no image/video) and always reasons before answering (thinking depth adjustable, not fully disableable). Multimodal and advanced features are API-only.
  • Hardware reality: a 2.4T checkpoint needs multi-node GPU infrastructure. For local/single-server use, the smaller Qwen3.8-27B is the practical open-weights option.

For research reproducibility, data sovereignty, and fine-tuning, this is a big deal: it's the first time frontier-class capability ships in open weights from Alibaba.

What We Liked

  • Frontier performance for a bargain price — PaperBench 93.0, GPQA 92.6, #2 vision on Arena, at $2/$6 per M tokens
  • First open Max-class flagship — Qwen3.8-2.4T-A95B weights are public, a milestone for the open ecosystem
  • Huge 1M context — long documents, full codebases, and long-horizon agent traces fit in one window
  • Native multimodal (API) — text, image, and video input in one model
  • Demonstrated long-horizon autonomy — internally completed a real software engineering project over a 16-day run (500+ chip-design turns reported)
  • Same price as GPT-5.6, ~5× cheaper than Claude Fable 5 on input

What We Didn't Like

  • Unverified, self-reported benchmarks — no independent runs confirmed the headline scores at launch
  • High agentic cost — 64 turns and ~$1.14 per task in independent testing; per-task economics are not as cheap as the token price suggests
  • Hallucination risk jumped — measured rate rose from 23% to 40% in independent agent testing
  • Weak on HLE — 43.6 on Humanity's Last Exam, the lowest among flagship models
  • SWE-bench Pro trails — 67.7 vs Fable 5's 80.0; not the #1 coding model
  • Too big to self-host — 2.4T params demands multi-node datacenter hardware
  • Open version is text-only and always-reasoning, so the open weights don't fully match the API experience

Qwen3.8-Max vs Claude Fable 5 vs GPT-5.6 vs Kimi K3

DimensionQwen3.8-MaxClaude Fable 5GPT-5.6Kimi K3
Params2.4T (95B active)ClosedClosed2.8T (104B active)
Context~1M~1M1M1M
Price in/out$2 / $6$10 / $50$2 / $6$3 / $15
Open weights✅ Yes❌ No❌ No✅ Yes
PaperBench93.0 (self)Frontier (claimed)
SWE-bench Pro67.780.0
MultimodalText/Image/Video (API)YesYesYes
Best forOpen AI, value, long-contextTop coding & reliabilityBalanced coding agentFront-end coding

Bottom line: choose Fable 5 if you need the most reliable coding and can pay the premium; choose Qwen3.8-Max if you want near-frontier capability at ~1/5 the price — or need actual open weights you can control.

Who Should Use Qwen3.8-Max?

  • 🎯 Enterprises that want frontier AI on a budget — same price as GPT-5.6, far below Claude
  • 🎯 AI teams that need open weights — fine-tune, self-host, and keep data in-house
  • 🎯 Researchers — the first open frontier-scale model with 1M context
  • 🎯 Long-context workloads — legal docs, codebases, long agent traces
  • 🎯 Multimodal agent builders (API tier) — image/video understanding at scale

Skip it if you need the most reliable code generation (Fable 5 still leads), you can't operate multi-node infrastructure for self-hosting, or you're deploying at the edge (use Qwen3.8-27B instead).

Final Verdict: 8.2 / 10

CategoryScore
Value for money9.5 / 10
Reasoning8.5 / 10
Coding7.5 / 10
Open source9.0 / 10
Multimodal (API)8.0 / 10
Reliability / efficiency6.5 / 10

Qwen3.8-Max is the value king and the open-source milestone of 2026. It doesn't unseat Claude Fable 5 as the most capable or most reliable model — the SWE-bench gap and the hallucination spike prove that — but at $2/$6 per M tokens with open weights, no other frontier model delivers this much capability for this little money. The one thing to watch: verify the benchmarks yourself before you bet production workloads on it.

Frequently Asked Questions

Is Qwen3.8-Max free to use?

The API is paid — $2/$6 per million tokens — but the open weights are free to download (you pay only for the infrastructure to run them). Consumers get limited free access through Qwen apps and Alibaba's Qingbei (通义) products.

Is Qwen3.8-Max better than DeepSeek?

Alibaba claims Qwen3.8-Max outperforms DeepSeek's latest on most benchmark families, with a much larger parameter count and 1M context. But DeepSeek remains far cheaper for high-volume use, and both sets of figures are vendor-reported — head-to-head third-party testing is still pending.

Does Qwen3.8-Max support vision and video?

Yes — in the hosted API. Inputs include text, images, and video. The open-weight version is text-only.

When will Qwen3.8-Max be available in my region?

It's available immediately through Alibaba Cloud Model Studio and OpenRouter; international availability continues to expand. Smaller Qwen3.8-27B open weights are available globally via Hugging Face / ModelScope.

Sources

Related Reviews

Disclaimer: Benchmark figures are vendor-reported as of August 2026 and subject to change. This article is for informational purposes; verify current pricing and capabilities before making purchasing decisions.

没有评论:

发表评论

Microsoft MAI Review 2026: The 7-Model Family That Says Goodbye to OpenAI — Trillion-Parameter Flagship, Now Default in GitHub Copilot

Microsoft MAI Review 2026: The 7-Model Family That Says Goodbye to OpenAI — Trillion-Parameter Flagship, Now Default in GitHub ...