Qwen4-Max: 2.4T Params, 71B Activated, 58% Cost Drop — But Max Isn't Open Source
Alibaba's Qwen4-Max uses a 2.4T-parameter MoE architecture activating only 71B parameters per token, achieving 58% lower per-task costs with native vision and 1M context, but the flagship model remains closed-source while smaller open-weight versions are promised later.
What Was Released
On November 16, Alibaba officially launched Qwen4-Max , the first model in the Qwen4 series. It is built on a new QSA (Qwen Sparse Attention) MoE architecture with 2.4 trillion total parameters but only 71 billion activated parameters per token (approximately 3% activation ratio, 25% fewer activated parameters than the previous generation). Key capabilities include native vision, 1 million token context window, stable long-horizon task execution, and the ability to call professional tools with phased delivery (called LiveDelivery). The API is live on the Qwen AI platform with Token Plan access, and the model is also available via QwenCloud, Qwen Office, Qoder, Qwen App, and Qwen Studio.
Critically, Qwen4-Max will not be open-sourced . The official announcement states: "The open weights of smaller Qwen4 models will be released. Stay tuned." Only mid- and small-sized models are slated for open release.
Technical Core: What 2.4T Total / 71B Activated Means
The activation ratio is the key lever. With 2.4T total parameters but only 71B activated per token (~3%), the model keeps capacity for knowledge storage while drastically reducing inference compute. In MoE terms, total parameters determine "whether it fits," while activated parameters determine "how fast and cheap it runs." Pushing activation down to 71B cuts the inference bill while preserving capability.
Combined with QSA sparse attention , the KV cache and attention overhead for long contexts are also reduced — this underpins the claimed 1M context with "cost down 60%."
Benchmarks (Self-Reported)
Terminal-Bench 4.0 (Xhigh tier): Average tokens per attempt dropped from 289K to 151K (~50% reduction); score rose from 37.9 to 48.5 (some media cite 38.9 as baseline).
Work360 (Alibaba internal benchmark): Score 47.74 → 52.69 ; average per-task cost 26.58 CNY → 11.16 CNY .
Third-party leaderboards: CodeArena frontend coding blind test global #6 (reported between GPT6-Astra and Claude Fable5.1); ALE 46.7, Chartography 54.2, JobBench 67.3.
All above scores are from Alibaba's internal evaluations unless noted.
Pricing: The Real Killer Is Cache Hits
Domestic pricing per million tokens: input 10 CNY, output 30 CNY, cache hit 0.5 CNY ; international: $1.5 / $4.5 / $0.075 respectively.
The standout is the cache hit price of 0.5 CNY . In long-horizon tasks, the vast majority of tokens are repeated context reads; high cache hit rates drive a cliff-like bill reduction. This cache pricing contributes the largest share of the "58% per-task cost drop," more aggressive than simple token price cuts.
Official case studies (not independently verified): automotive workpiece programming manual effort reduced from ~475 minutes to 24–26 minutes (~95% reduction); AI4S scenario ran 48 hours continuously, screening 2.7 million candidate molecules to discover a novel coolant molecule.
Honest Boundaries: Don't Be Misled by "Down 60%"
Mostly self-reported numbers. Work360 is Alibaba's own framework with no third-party replication; CodeArena "global top 6" citations vary across outlets; native vision lacks public benchmarks. Developer retention depends on external validation.
Metric discrepancies. Baseline scores cited as 37.9 vs 38.9; programming time 24 vs 26 minutes; "58%" vs "60%" reduction. This article uses mainstream figures and notes differences, not treating them as exact facts.
Comparison baseline is Qwen3.8-Max, not the 11-11 Qwen4-Max-Preview. Official notes post-training affects comparability.
Max is not open-source. 2.4T total parameters with FP4 weights likely remain in the terabyte range — cannot fit in a 512GB Mac. Local deployment enthusiasts must wait for the upcoming mid/small open-weight releases.
Decision Guide: Should Your Mac Wait?
Local-only Mac users: Don't rush to switch. Current Qwen3.8 / GLM / Llama run fine. Qwen4-Max is irrelevant locally; watch for the "successively open-sourced" mid-size models — likely ~27B dense variants that fit 32GB/64GB Macs. Wait, but don't pause your work.
Existing Token Plan / heavy cloud API users: Switch directly. Same price buys more tokens, cheaper cache, more stable long tasks — no reason not to migrate.
Casual / occasional users: Try free/low-cost entry points first (Qwen App, Qoder, Qwen Studio). Price war just started; wait for better deals.
Enterprise / professional toolchains (3D, finance, legal): Evaluate rigorously. Official pitch emphasizes "call professional tools + phased delivery," but must test with your own real tasks , not trust benchmark scores.
Author's Take
This release shifts LLM competition toward real productivity per unit cost — the parameter arms race is pausing; whoever makes a task cheaper to complete wins developers. Cutting cache price to 0.5 CNY is more practical than chasing leaderboard points.
For local deployers: Max is just an architectural direction signal; the small open models are your meal. Once QSA sparse attention + ultra-low activation ratio trickles down to ~27B, Qwen4 on a 32GB Mac will be highly promising — but that's future. Today, keep running your machine as you do.
Source
Primary source: Qwen official tweet (@Alibaba_Qwen, 2026-11-16) announcing Qwen4-Max general availability, API on QwenCloud, and preview of smaller open-source Qwen4 models. Secondary coverage: Sina Finance, NetEase Tech, Tencent News, AIBase, Pacific Tech (all 11-16). Specs and scores per Alibaba official disclosure; third-party replication pending; prices as of launch day, subject to change. This article is technical interpretation, not purchase advice.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
