Back to Blog

Qwen4-Max: 2.4T Parameters, 1M Context and Lower Costs

Daily News6436
Qwen4-Max: 2.4T Parameters, 1M Context and Lower Costs

Alibaba's Qwen team formally launched Qwen4-Max on November 16, the newest flagship in the Qwen4 series. The model moves past the preview stage that opened on November 11 and is now the production-grade entry into the Qwen4 API lineup. Developers can reach it through the Token Plan or the standard API surface.

The release is defined by three numbers: 2.4 trillion total parameters, a native context window of one million tokens, and a per-task cost reduction exceeding 60% against its predecessor. Each carries a different implication for enterprise buyers, and none of them is fully explained by the launch material.

Architecture: Sparse MoE With Three Structural Changes

Qwen4-Max uses a sparse Mixture-of-Experts backbone. Above that base, the team reports systematic changes in three dimensions — attention, residual connection, and embedding.

Gated DeltaNet in the attention stack. Three of every four layers compress historical state into a fixed-size representation using Gated DeltaNet. The remaining layer applies Qwen Sparse Attention to retrieve relevant context at block granularity. The combination targets the dominant cost of long-sequence inference: quadratic attention over a growing key-value cache. Fixed-state compression bounds memory growth. Block-level sparse retrieval restores precision where it matters.

Gated Residual with four parallel branches. The residual stream is split into four parallel paths, which the team describes as improving gradient propagation across depth. In practice this is a stability intervention — very deep, very wide sparse models are prone to training instability, and widening the residual channel is one way to mitigate it.

N-gram Embedding at the input layer. A histogram-style representation supplements token embeddings before the first layer. The stated goal is capacity expansion at low incremental compute cost. Input-side enrichment is cheaper than adding depth or width, so the tradeoff is favorable when the vocabulary and phrase distribution are stable.

The claimed outcome of these changes is a reduction in both time-per-token and cost-per-task relative to Qwen3.8-Max. The team has not published the absolute latency figures behind that claim.

Benchmark Position: Sixth on Code Arena WebDev

The headline external result is a sixth-place ranking on Code Arena's WebDev leaderboard under the xHigh reasoning setting, with a score of 1,750. Qwen4-Max lands above several closed frontier models on that specific board.

RankModelScore
1Claude Fable 5.2 (xHigh)1,826
2Claude Opus 5.5 (Max)1,816
3Claude Sonnet 5.5 (xHigh)1,772
4Gemini 4 Carbon (High)1,768
5GPT-6 Astra (Max)1,763
6Qwen4-Max (xHigh)1,750
7Claude Fable 5.1 (Max)1,743
8GPT-6.1 Sol (Max)1,742

Two caveats apply. First, Code Arena is a preference-based arena, not a correctness benchmark; rankings over a 100-point band reflect human judgment on web-development output quality, and the gaps here are within noise range of one another. Second, five of the models above Qwen4-Max are running at maximum reasoning effort while Qwen4-Max is at xHigh. The comparison is close but not clean.

Qwen's own internal evaluation reports improvement over Qwen3.8-Max across all tested dimensions, spanning productivity, creative writing, software development, and scientific problem solving. Those figures are vendor-reported and not independently replicated.

The Efficiency Numbers Are the Real Story

The cost data carries more decision weight than the leaderboard placement. Qwen reports the following against Qwen3.8-Max at matched reasoning tiers:

MetricChange
Token consumption, xHigh tier−24%
Token consumption, High tier−35%
Terminal-Bench 4.0 tokens (high reasoning)−40%
Terminal-Bench 4.0 tokens (standard)−49%
API price~50% lower
Work360 cost per task¥26.58 → ¥11.76 (−58%)

The compounding effect matters. Token reduction and unit price reduction multiply rather than add. A 35% token cut on top of a 50% price cut yields a per-task cost well under half the prior figure, which is how the 58% Work360 reduction is derived.

For agent workloads, token volume is the dominant cost driver, because each task may issue dozens of calls. A model that consumes a third fewer tokens per call changes the economics of retry-heavy pipelines more than a headline price cut does.

Work360: A Vendor-Built Benchmark for Real Office Work

Qwen's team built its own evaluation harness rather than relying solely on public leaderboards. Work360 spans 26 top-level industries, 182 secondary domains, and 21 professional directions, drawing task definitions from actual industry workflows rather than synthetic prompts. Qwen4-Max scores 52.69 on it against Qwen3.8-Max's 47.74 — a 4.95-point gain.

Self-designed benchmarks warrant scrutiny. A benchmark authored by the model's own vendor, on tasks selected by that vendor, tends to reward the capabilities that vendor has prioritized. Work360 is useful as a directional signal and as a description of where Qwen believes its model is strong. It is not a neutral comparison against Claude or Gemini.

That said, the domain coverage is more granular than most public agent suites, and the inclusion of legal review, financial research, and clinical documentation alongside engineering tasks is a reasonable proxy for enterprise demand.

Multimodal Input and the Live Delivery Mechanism

Qwen4-Max accepts native multimodal input, including images and video, and carries the full one-million-token context window into those modalities. The practical effect is that a long video, a large slide deck, or a scanned document set can be processed without an intervening transcription or extraction stage.

The release also introduces Live Delivery. The default behavior returns a first chunk quickly, then streams subsequent content based on user feedback — a latency-optimization pattern aimed at perceived responsiveness rather than throughput. A companion Qwen Live Delivery Plugin is shipping alongside it. The mechanism reduces time-to-first-output; it does not reduce total generation time, so teams should not conflate the two in latency budgeting.

Reported demonstration cases span Blender-based 3D modeling, Unity and Unreal game development pipelines, automotive parts modeling with machining program generation, financial research, legal and regulatory review, and clinical documentation. In the automotive parts case, single-workpiece programming time dropped from roughly 4.5 minutes to 24 seconds. That figure is a vendor demonstration under controlled conditions and should be treated as an upper-bound illustration, not a production expectation.

Availability and Roadmap

Qwen4-Max is live on QwenCloud for API access. It is also reachable through Qwen Studio, Qwen Code, and the consumer-facing web product. Qwen positions Qwen4-Max as the flagship of the Qwen4 series and states that smaller open-source variants will follow along the same technical路线 — that is, the same sparse-attention and compression design will be scaled down rather than replaced.

That roadmap signal is worth noting for teams with data-residency constraints. If the architectural approach survives downscaling, on-premise deployment of a mid-sized Qwen4 variant becomes a realistic option where a 2.4-trillion-parameter API-only model is not.

What Remains Unconfirmed

Several items required for procurement decisions have not been published:

The last item is the one to watch. Vendor benchmarks and arena preference scores both measure something real, but neither substitutes for a task-specific evaluation on the buyer's own workload.

Teams running mixed-vendor model comparisons typically need a consistent request surface and unified credential handling across providers, which is where an API gateway layer earns its place in the stack. 4sapi provides that style of unified model access, letting a single integration test Qwen4-Max against Claude and Gemini endpoints without rebuilding client code per vendor.

Bottom Line

Qwen4-Max is a credible frontier-tier release on cost efficiency. The 2.4-trillion-parameter sparse architecture, the million-token native context, and the combined token-and-price reduction give it a defensible position in agent-heavy and long-document workloads. Its external benchmark standing is competitive but not dominant, and the most favorable published numbers come from vendor-authored evaluations.

The decision-relevant question is narrower than the leaderboard: for a given task family, does a 35–49% token reduction at half the unit price offset a capability gap of a few arena points? For high-volume, retry-heavy pipelines, the arithmetic usually favors the cheaper model. For tasks where a single incorrect output carries material cost, it usually does not.

Tags:Qwen4-MaxAlibaba QwenQwen4 API2.4T ParametersMillion Token ContextSparse MoEGated DeltaNet

Recommended reading

Explore more frontier insights and industry know-how.