Qwen3.8-Max: Alibaba's 2.4T Open-Weight Coding Model

Qwen3.8-Max is Alibaba's newest flagship large language model — a 2.4-trillion-parameter Mixture-of-Experts (MoE) model the Qwen team calls its "most capable model to date." It's pitched squarely at coding and "cowork": long-horizon, autonomous work where an agent plans, executes, and self-corrects with little hand-holding. The headline for builders is that the weights are going open. Here's what Qwen3.8-Max actually is, what it costs, what's genuinely confirmed versus still unclear, and where it fits in a real agent stack.

What Is Qwen3.8-Max?

Qwen3.8-Max is a 2.4-trillion-parameter sparse MoE model from Alibaba's Qwen team, positioned as the most powerful entry in the Qwen family. Alibaba has made it broadly available through its hosted surfaces and confirmed that open weights ship shortly after launch, alongside a second, much smaller open-weight checkpoint — Qwen3.8-27B — aimed at teams that want to run a model on their own hardware. (MarkTechPost)

Two things make the release notable. First, if the weights ship as promised, this would be the first Max-class Qwen model Alibaba has ever released openly — the flagship tier, not a distilled sibling. (How2Shout) Second, the model is multimodal by design: it accepts text, image, and video as input and returns text, with vision framed as part of the execution loop rather than a one-way input channel. (MarkTechPost)

Specs at a Glance

Based on Alibaba's announcement and early launch coverage, here's what's on the table:

• 2.4T total parameters, sparse MoE. That headline number would make Qwen3.8-Max one of the largest open-weight models announced to date — larger than DeepSeek V4 Pro (~1.6T), though smaller than Kimi K3 at 2.8T. (techsy)
• ~1M-token context reported in launch coverage — enough to hold a full codebase or a large document corpus in view. (How2Shout)
• Native multimodal input — text, image, and video in; text out. (MarkTechPost)
• OpenAI- and DashScope-compatible API, so wiring it into an existing stack is largely a base-URL and model-ID change. (MarkTechPost)
• Qwen3.8-27B — a second checkpoint also going open-weight, and the one that realistically fits ordinary on-premise GPU hardware. (How2Shout)

One important gap: Alibaba has not disclosed the activated-parameter count — how many of the 2.4T parameters actually fire per token. In a sparse MoE, that number, not the headline total, drives serving cost. Without it, inference cost for the open weights can't yet be modeled. (MarkTechPost)

The Capabilities Alibaba Is Selling

The Qwen team framed Qwen3.8-Max around long-horizon, autonomous work rather than single-shot chat. The headline demos, per Alibaba's launch materials: (latent.space)

• Multi-day autonomous coding — a self-evolving development run that went from an empty folder toward a working project over 10-plus days, with the full trace published on GitHub.
• Long-horizon planning — system-level autonomous planning with closed-loop learning, illustrated by 500-plus turns of chip-design optimization and a 365-day e-commerce strategy simulation.
• Autonomous research — Alibaba describes rebuilding a research paper's pipeline from scratch and iterating on it over roughly 125 hours to produce a data-selection method that beat the original paper's benchmark by a small margin. (latent.space)
• Multimodal feedback loops — vision used continuously for planning, execution, and self-correction, not just as an input. (MarkTechPost)

These are vendor claims from launch demos. They're directionally interesting — the whole industry is pushing toward agents that run for hours or days unattended — but they aren't yet third-party-verified benchmarks. Treat them as a starting point, not a settled verdict.

Pricing

Alibaba published API pricing alongside the launch: (latent.space)

That undercuts most U.S. frontier models by a wide margin and sits below Kimi K3's hosted rate (~$3 input / $15 output), continuing the pattern of Chinese labs competing hard on price. The cheap cached-input rate rewards workloads that reuse a large fixed context — exactly the shape of many coding and document agents.

What's Confirmed vs. What Isn't

This launch moved fast, and coverage is uneven, so it's worth separating the two.

Reasonably confirmed:

• 2.4T total parameters and a plan to open the weights, per Alibaba's own announcement. (techsy)
• API access is live today and is OpenAI/DashScope-compatible. (MarkTechPost)
• Published API pricing of $2 / $6 / $0.25 per million tokens. (latent.space)
• A second open-weight checkpoint, Qwen3.8-27B. (How2Shout)

Not yet settled:

• The activated-parameter count and full MoE configuration, so serving cost is unknown. (MarkTechPost)
• A complete, independent benchmark table. Early reporting noted the launch materials lacked a full benchmark set at the time of writing; check for third-party results before betting a workflow on it. (MarkTechPost)
• The exact license and whether the released checkpoint matches the preview. "Open weights" typically means downloadable weights, not necessarily training data, training code, or an OSI-approved license. (coursiv)

The honest read: Qwen3.8-Max is real and available, the open-weight promise is on record, but the numbers that would let you plan production capacity aren't all public yet.

Can You Actually Run It?

For most teams, the 2.4T flagship is aspirational to self-host. A 2.4-trillion-parameter checkpoint is a multi-node datacenter artifact — no individual developer runs it on a workstation, and many companies can't host it either. (How2Shout)

That's why Qwen3.8-27B is the more practical story. It's the checkpoint sized for ordinary on-premise GPUs, and for most self-hosting scenarios it — not the flagship — is the realistic deployment path. (MarkTechPost) If your goal is data residency or cost control on your own hardware, plan around the 27B; if you want the flagship's ceiling, the hosted API is the near-term route.

How Qwen3.8-Max Fits the Open-Weight Race

Qwen3.8-Max lands in a crowded few weeks for large open-weight models. It follows Kimi K3 (2.8T) from Moonshot AI and Thinking Machines' Inkling, and it sits above DeepSeek V4 Pro on raw parameter count. (techsy) The competitive signal is clear: frontier-scale open weights are arriving on a roughly monthly cadence, and the gap between open and closed models keeps narrowing.

For builders, the practical takeaway isn't picking a single "winner." It's that you now have several credible, cost-effective, self-hostable-in-principle frontier models — and the leverage comes from routing each task to the right one.

Put a Model Like Qwen3.8-Max to Work in Your Own AI Workforce

A frontier model earns its keep when it's wired into real work — repos, documents, terminals, and multi-step plans — not just a chat box. That's what a model-agnostic, multi-agent platform is for: route long-horizon coding to a model like Qwen3.8-Max, keep cheaper models for routine steps, and orchestrate them across actual tasks. Eigent is an open-source Cowork desktop app that runs a multi-agent AI workforce locally, so you can plug in the best model per task and keep sensitive data on your own machine. See how agents can review GitHub PRs end to end, then download Eigent to try it.

Frequently Asked Questions

What is Qwen3.8-Max?

Qwen3.8-Max is Alibaba's flagship large language model — a 2.4-trillion-parameter sparse Mixture-of-Experts model built for coding and long-horizon agentic work. Alibaba calls it its most capable model to date and has committed to releasing the weights openly, alongside a smaller Qwen3.8-27B checkpoint.

Is Qwen3.8-Max open source?

Alibaba has said the weights for Qwen3.8-Max and Qwen3.8-27B will be released openly. "Open weights" generally means the weights are downloadable — not necessarily the training data, training code, or an OSI-approved license. The exact license had not been fully detailed at launch, so confirm the terms before relying on redistribution rights.

How much does Qwen3.8-Max cost?

Alibaba's published API pricing is $2.00 per million input tokens, $6.00 per million output tokens, and $0.25 per million cached-input tokens — well below most U.S. frontier models.

Can I run Qwen3.8-Max myself?

The 2.4T flagship is a multi-node datacenter model that most individuals and companies can't self-host. The smaller Qwen3.8-27B checkpoint is the realistic on-premise path for ordinary GPU hardware; for the flagship, the hosted API is the near-term option.

How does Qwen3.8-Max compare to Kimi K3?

On raw parameter count, Qwen3.8-Max (2.4T) is smaller than Kimi K3 (2.8T) but larger than DeepSeek V4 Pro (~1.6T). Its API pricing is lower than K3's hosted rate. Head-to-head quality comparisons will have to wait for independent benchmarks, which were not fully published at launch.

Can I use Qwen3.8-Max with a multi-agent platform like Eigent?

Yes. The API is OpenAI- and DashScope-compatible, so a model-agnostic platform like Eigent can route coding and knowledge-work tasks to Qwen3.8-Max while using other models for routine steps — and keep data local when needed.