Qwen3.8-Max

TL;DR. Qwen3.8-Max (Alibaba) is the first Qwen Max tier shipped as weights. Strengths — a 2.4T total / 95B active MoE that, on author-reported figures, hits PaperBench 93.0 ahead of GPT-5.6 Sol and Opus 4.8, and Terminal-Bench 2.1 86.6 past Opus 4.8 and Fable 5 (84.6). Weaknesses — the published checkpoint is not the API model: the HF card states text-only with 262,144 native context, and the license is a house qwen3.8-max agreement rather than a standard OSI one. SWE-bench Pro at 67.7 falls well short of Fable 5 (80.0), and there is no independent composite index listing yet.

Provenance note. This is not a hands-on review of Qwen3.8-Max. The Hugging Face model card was read directly to confirm license, context, and modality (primary); the benchmark figures come from third-party coverage relaying the announcement and are marked as author-reported (secondary). The firmest facts in this piece are the card's specs, not the benchmarks. No measured or felt observations.

At a glance

Why now — the Max tier opened

Qwen ran for a long time on a pattern of opening the small and mid tiers while keeping Max closed. Qwen3.8-Max is the first model past that line — a 2.4T flagship released as weights, three weeks after Kimi K3 (2026-07-27, 2.8T).

Two labs throwing trillion-parameter flagships open within a month is the signal for this category. The framing of open weights as "the small, cheap alternative" ends here.

Strengths — what it does well

1. Max tier opened — a first for the line

This is the first time a Max-class Qwen is downloadable. The official repo is Qwen/Qwen3.8-2.4T-A95B in BF16 safetensors, compatible with vLLM and SGLang. That one more top-band model can be held directly matters more than the benchmark figures.

2. Overlaps the closed frontier on agentic and research benchmarks (author-reported)

Per the announcement, PaperBench 93.0 leads GPT-5.6 Sol and Opus 4.8, and Terminal-Bench 2.1 86.6 passes Opus 4.8 and Fable 5 (84.6) while trailing GPT-5.6 Sol max (88.8). On research-document handling and terminal agents it lands on the same table as the closed frontier. All vendor figures, so read direction only until a standardized board.

3. A small checkpoint was targeted alongside it

Alibaba previewed Qwen3.8-27B, aimed at ordinary on-premise GPU hardware, for the same week. Pairing a trillion-parameter flagship with a 27B in one generation is Qwen's own axis in this category — but as of 2026-08-15 the 27B has not shipped (no repo, card, or license file). Do not put the preview into an adoption plan.

Weaknesses — what it does poorly

1. The published checkpoint is not the same artifact as the API model

This is the most important fact in the piece, and the only weakness confirmed against a primary source. The announcement and coverage introduced Qwen3.8-Max with text, image, and video input at 1M context. The actually published Qwen/Qwen3.8-2.4T-A95B card states it is text-only — "Multimodal inputs are not supported." Context is listed as 262,144 native, with 1M as an extended figure.

So the model you picked off a benchmark is not the model you downloaded. Expect multimodality from the weights and it is not there. This model is the case that made card-versus-API-spec reconciliation a required step in this category.

2. The license is not standard OSI

The HF card's license is a house agreement named qwen3.8-max. It is not a settled artifact like GLM 5.2's MIT, and it is a separate instrument from the Kimi K3 License too. Commercial terms have to be read directly and put through legal review. Some coverage reports a revenue-sharing requirement for large commercial users, but that is not confirmed in the model card body — check the clause text itself.

3. Core coding benchmarks fall short of the closed frontier

Even on author-reported figures, SWE-bench Pro is 67.7 against Fable 5's 80.0 and FrontierSWE 73.5 against 88.8. Unlike the agentic and document axes in strength 2, the gap on core code modification is clear. The "beats Opus 4.8" headline rests on one Terminal-Bench cell, not the whole field.

4. No independent composite index listing

Kimi K3 (57) and GLM 5.2 (51) are listed on the AA Intelligence Index and so compare on one axis. As of 2026-08-15 that slot is empty for Qwen3.8-Max, so it cannot be ranked within this category. Ranking it off vendor figures alone would be wrong.

5. Thinking mode is mandatory

The card states that thinking mode is required for all interactions. On short, repetitive calls those thinking tokens accumulate straight into cost and latency. It is structurally a poor fit for high-volume low-latency workloads.

What to use it for

• You must hold Max-class weights directly → Qwen3.8-Max. Read the license clauses first.
• Research documents and terminal agents → Qwen3.8-Max (PaperBench, Terminal-Bench; author-reported). Recheck after a standardized board.
• You need multimodality → not from the published checkpoint (text-only). Kimi K3 holds native vision among open weights.
• Minimize licensing risk → GLM 5.2 (MIT).
• Top capability in open weights → Kimi K3 (AA 57). Qwen has no independent index listing at all.

Conclusion

The news in Qwen3.8-Max is that the Max tier opened; the lesson is that "opened" has to be verified against the spec.

1. Scope of the opening — you can obtain a Max-class 2.4T. A first for the line, and it widened this category's ceiling.
2. Terms of the opening — the license is a house agreement with review outstanding, and the published checkpoint is narrower than the API model (text-only, 262K native). Those two are the substance of the adoption call.
3. Capability — the agentic and document axes overlap the closed frontier but are author-reported, and core code modification still trails. Hold any ranking claim until an independent index listing.

Sources

Primary

• Qwen/Qwen3.8-2.4T-A95B model card (Hugging Face) — license, context, modality, thinking-mode requirement (the only primary source verified directly for this piece)

Announcement and coverage (author-reported figures)

• MarkTechPost — Alibaba Qwen releases Qwen3.8-Max (architecture, announced benchmarks)
• DataCamp — Qwen3.8-Max features, benchmarks, and pricing

Related

• Alternative Models overview — ranked by the terms of the opening
• Kimi K3 — capability leader among open weights
• GLM 5.2 — MIT and the cheapest slot
• models overview — ranking base models