Alibaba

Qwen3.8-Max

Provider

Alibaba

Status

Current

Context

1,000,000 tok

SWE-bench

67.7%

Price

$2 / $6 /MTok

Qwen3.8-Max is Alibaba’s new flagship Qwen model, released on 3 August 2026 after a preview on 19 July. It is a ~2.4-trillion-parameter mixture-of-experts reasoning model with a 1-million-token context window and, for the first time in the Max line, multimodal input (text, image and video in; text out). Alibaba positions it as its most capable model yet — claiming it is “second only to Fable 5” and beats GPT-5.6 Sol and Claude Fable 5 on several of its cited coding, general and multimodal benchmarks.

Two things make the launch notable beyond the benchmarks. First, price: at $2 input / $6 output per million tokens it landed three days after OpenAI cut its GPT-5.6 Terra tier to $2/$12, deliberately halving OpenAI’s output price at a similar capability tier. Second, open weights: in August Alibaba published a text-only open checkpoint of the full 2.4T model, Qwen3.8-2.4T-A95B, under its own Qwen3.8-Max License, and Qwen3.8-27B under Apache 2.0, making this the first Qwen Max-class model with open weights. Artificial Analysis scores the hosted 0902 snapshot 45 on its Intelligence Index v4.3.2, level with GLM-5.3 and a point above Kimi K3, at $5.41 per completed task, more than twice their cost.

Quick specs

VIEW QWEN3.8-MAX →

The 0902 snapshot (2 September 2026)

Alibaba released Qwen3.8-Max-0902 on 2 September 2026 — an upgraded snapshot of the same model, post-trained for coding and agentic work, keeping the version number, the ~2.4T parameters, the 1M-token context and the $2/$6 pricing unchanged. The measurable claim: its front-end CodeArena score rose 22 points to 1,691, taking first place on that leaderboard (TechNode). CodeArena is a public arena rather than a vendor table, which is more than most of this page’s numbers can say — and Artificial Analysis has since scored the 0902 snapshot 45 on its re-based Intelligence Index v4.3.2.

What Qwen3.8-Max is

Qwen3.8-Max is a reasoning-native, agentic flagship. It ships with an explicit thinking mode (reasoning budget up to ~262K tokens), function calling, structured outputs and tool use, and it targets long-horizon agentic work — Alibaba routes it through its Qoder agentic IDE and exposes OpenAI-, DashScope- and Anthropic-compatible endpoints so external harnesses like Claude Code, Cursor, Cline and Codex work by swapping the base URL and model ID.

Architecturally it is a mixture-of-experts model of about 2.4 trillion total parameters — the first Qwen Max-tier model confirmed as MoE and above a trillion parameters. The open checkpoint’s name and model card confirm 95 billion active parameters per token. The 1-million-token context holds roughly 991K input tokens (about 983K with thinking on) and up to ~131K output. Unlike the text-only Qwen3.7-Max, this generation accepts image and video input alongside text, making it Qwen’s first multimodal model above a trillion parameters. The knowledge cutoff is not disclosed.

The headline strategic move is the reversal on openness. Where Qwen3.7-Max broke with tradition by shipping closed and API-only, Alibaba says Qwen3.8-Max will return to open weights — the full 2.4T model plus a smaller Qwen3.8-27B aimed at on-prem GPUs — and it did, in August. The open 2.4T checkpoint on Hugging Face differs from the hosted model: it is text-only, requires thinking mode on every request, and has a 262K native context (extensible to about 1M), where the hosted Max adds image and video input, a non-thinking mode, built-in tools and a 1M default context. It carries the Qwen3.8-Max License: products above 100 million monthly users or $20 million monthly revenue must display the model name, and companies with more than $50 million in annual revenue need a separate commercial licence to offer it as a model-as-a-service or AI work-assistant business. Qwen3.8-27B is Apache 2.0, takes text, image and video input and has a 262K native context. The launch is widely read as Alibaba’s answer to Moonshot’s Kimi K3 and the broader Chinese open-weight race.

Benchmark performance

Alibaba published a broad vendor benchmark table; read it as a ceiling. The independent reading is Artificial Analysis’s: 45 on the Intelligence Index v4.3.2 for the 0902 snapshot, at $5.41 per task. There is still no LMArena Elo and no standardized (Scale) SWE-bench figure.

The pattern is a frontier-adjacent agentic and coding model on vendor numbers, priced far below the US frontier. The honest caution is precedent: Qwen3.7-Max also looked flagship-tier on Alibaba’s own benchmarks (80.4% SWE-bench Verified) yet landed mid-pack (index ~46, ~#12) once Artificial Analysis ran it independently — and AA flagged it as verbose, which eroded its price advantage. The independent score of 45 puts it level with GLM-5.3 and a point above Kimi K3, below the vendor table’s “second only to Fable 5” claim, and Artificial Analysis measured it as verbose (190 million output tokens to run the index), which is why a completed task costs $5.41 despite the low token price. See best AI models for the full ranking.

Pricing and access

Qwen3.8-Max is priced at $2.00 input / $6.00 output per million tokens on Alibaba Cloud (Token Plan), with implicit cache reads at $0.25 — aggressive for a model of this class, and a deliberate shot at OpenAI, which had cut its GPT-5.6 Terra tier to $2/$12 three days earlier. Qwen matched the input price and halved the output price at a comparable capability tier. Preview access ran at up to ~90% off the standard rate.

The API model ID is qwen3.8-max, reachable through the international endpoint (dashscope-intl.aliyuncs.com) or the China endpoint, with OpenAI-, DashScope- and Anthropic-compatible modes. It is also wired into Alibaba’s Qoder and QoderWork agentic tools and the consumer Qwen app.

Open weights shipped in August. The text-only 2.4T checkpoint is on Hugging Face under the Qwen3.8-Max License, and Qwen3.8-27B is on Hugging Face under Apache 2.0; Artificial Analysis prices hosted Qwen3.8-27B at $0.50/$3.00 per million tokens. Hosted Qwen3.8-Max routes through Alibaba Cloud (China jurisdiction).

How Qwen3.8-Max compares

• vs the proprietary frontier — On vendor benchmarks it lands just under Claude Fable 5 and around GPT-5.6 Sol, at roughly a third to a half of their price. But Sol, Opus 5 and their peers carry independent Artificial Analysis and arena results; Qwen3.8-Max now has one too, 45 on Artificial Analysis’s index, which places it below the flagship pack and level with GLM-5.3.
• vs other Chinese models — It is the direct answer to Moonshot’s Kimi K3, which reached 4th on the independent intelligence index and shipped open weights on 26 July. On Artificial Analysis’s re-based index Qwen3.8-Max (45) sits a point above Kimi K3 (44) and level with GLM-5.3 (45), but at more than twice their cost per completed task ($5.41 against about $2).
• vs its predecessor Qwen3.7-Max — cheaper ($2/$6 vs $2.50/$7.50), multimodal rather than text-only, stronger vendor coding and agentic scores, and — crucially — returning to open weights rather than staying closed.

Known limitations

Costly per task — Artificial Analysis measures $5.41 per completed index task, more than twice GLM-5.3 or Kimi K3, because the model is verbose, and it runs at 39 output tokens a second. The open 2.4T checkpoint is not the hosted model — it is text-only, always reasons, and its licence requires a separate commercial licence for model-as-a-service and AI work-assistant businesses above $50 million in annual revenue. China-hosted — hosted access carries data-residency, compliance and content-control considerations. Undisclosed internals — knowledge cutoff, SWE-bench Verified and MMLU/AIME figures were not published.

Frequently asked questions

What is Qwen3.8-Max?

Qwen3.8-Max is Alibaba’s flagship Qwen model, released 3 August 2026 — a ~2.4-trillion-parameter mixture-of-experts reasoning model with a 1-million-token context window and multimodal (text, image, video) input, built for long-horizon agentic and coding work.

Is Qwen3.8-Max open source?

Partly. The hosted Qwen3.8-Max is proprietary, but in August 2026 Alibaba published a text-only open checkpoint of the same 2.4T model, Qwen3.8-2.4T-A95B, under the Qwen3.8-Max License, which requires a separate commercial licence for model-as-a-service and AI work-assistant businesses with more than $50 million in annual revenue. Qwen3.8-27B is fully open under Apache 2.0.

How much does Qwen3.8-Max cost?

$2.00 per million input tokens and $6.00 per million output on Alibaba Cloud, with cached input at $0.25. That deliberately undercut OpenAI, which had cut its GPT-5.6 Terra tier to $2/$12 three days earlier — Qwen matched the input price and halved the output price.

How good is Qwen3.8-Max?

Artificial Analysis scores the 0902 snapshot 45 on its Intelligence Index v4.3.2, level with GLM-5.3 and a point above Kimi K3, at $5.41 per completed task. Alibaba’s own figures are higher — 67.7% SWE-bench Pro, 92.6 GPQA Diamond and 86.6 Terminal-Bench 2.1, which it calls “second only to Fable 5” — and read as a ceiling.

What is Qwen3.8-27B?

Qwen3.8-27B is the open-weight 27-billion-parameter sibling, released in August 2026 under Apache 2.0. It takes text, image and video input with a 262K native context, scores 34 on Artificial Analysis’s Intelligence Index v4.3.2 at $1.01 per task, and Alibaba reports 61.7% on SWE-bench Pro.

How does Qwen3.8-Max compare to Qwen3.7-Max?

Qwen3.8-Max is cheaper ($2/$6 vs $2.50/$7.50), adds image and video input (the Max line was previously text-only), posts stronger vendor coding and agentic scores, and returned to open weights — where Qwen3.7-Max shipped closed and API-only.

Last verified 30 September 2026. The 3 August launch, 1M context, multimodal input and $2/$6 pricing are drawn from Alibaba’s launch materials and independent write-ups. The open checkpoints, their licences and the 95B active-parameter count are confirmed against Alibaba’s Hugging Face repositories. The Intelligence Index scores and cost per task are Artificial Analysis’s independent measurements; the benchmark table is vendor-reported (Alibaba’s own testing).