QWEN3.8-MAX — CONFIRMED vs UNCONFIRMED (AS OF AUGUST 3, 2026)
● ✓ CONFIRMED: 2.4T parameter sparse MoE — from Alibaba's official July 19 announcement
● ✓ CONFIRMED: Multimodal — text, images, video, documents
● ✓ CONFIRMED: Always-on reasoning, "continuously evolving" preview
● ✓ CONFIRMED: Available now via Token Plan, Qoder, QoderWork at ~10% of standard price
● ✓ CONFIRMED: Open weights promised "soon" — no date, license, or HuggingFace repo
● ✗ NOT CONFIRMED: $2/$6/M per-token API pricing — no standard API price published
● ✗ NOT CONFIRMED: Active parameter count — 2.4T is total, not active
● ✗ NOT CONFIRMED: Any published benchmark table or third-party evaluation
● ✗ NOT CONFIRMED: A smaller 27B variant — this is a confusion with older Qwen3 models
● ✗ NOT CONFIRMED: General release date or open-weight timeline
What Actually Happened — and When
Alibaba previewed Qwen3.8-Max on July 19, 2026, at the World Artificial Intelligence Conference in Shanghai. According to MarkTechPost's July 19 report, the announcement arrived two days after Moonshot released Kimi K3's 2.8T weights — making it the second major Chinese frontier model announcement in the same week. What went live was the preview endpoint: qwen3.8-max-preview, accessible through Alibaba's Token Plan subscription, Qoder, and QoderWork. The endpoint is described by the Qwen team as "continuously evolving" — preview language, not a production launch.
As eesel AI's hands-on review documents, the preview runs at approximately 10% of standard pricing via the Token Plan credit system — which starts at around $6/month for Lite tier access. There is no published standalone per-token API price. The $2/$6/M figure circulating in some X posts and AI news summaries is not confirmed by any official Alibaba source. For reference, the predecessor Qwen3.7-Max is priced at $2.50/$7.50/M with a 90% cached-input discount — any expectation for 3.8-Max pricing should start from that baseline, not from unconfirmed X posts.
The Performance Claim — Alibaba's Own Words, No Third-Party Verification
Alibaba's official claim is that Qwen3.8-Max is "second only to Fable 5" among frontier models. As Tech-Now.io's analysis notes plainly, this is a "bold claim — and, as of this writing, one Alibaba has not backed with a single published benchmark score." No benchmark table exists. No methodology. No comparison configuration notes. No independent evaluation from Artificial Analysis, LMArena, or any third party has been published as of August 3. Alibaba's previous flagship, Qwen3.7-Max (May 2026), launched with a complete benchmark table including an 80.4% SWE-bench Verified score. Qwen3.8-Max launched with a tweet and a ranking claim. The two launches are not equivalent in terms of verifiable evidence.
According to emergent.sh's independent review, one hands-on reviewer placed Qwen3.8-Max-Preview "in the top tier alongside Fable 5, GPT-5.6 Sol, Grok 4.5, and Kimi K3, with speed as the main differentiator holding it back for daily-driver use." That is one person's qualitative assessment of a preview build — useful signal, not a benchmark result.
What's Missing That You Need Before Making Decisions
No benchmark table. Alibaba has not published SWE-bench, Terminal-Bench, GPQA, or any other standard evaluation score. "Second only to Fable 5" is positioning, not a result.
No active parameter count. 2.4T is total parameters. For a MoE model, the active parameter count (what fires per token) determines serving cost and latency. Alibaba has not disclosed it. As Techsy.io notes, "a sparse 2.4T model and a dense 2.4T model are very different things to serve."
No per-token API price. The Token Plan credit system does not publish a token conversion rate. The $2/$6/M figure in circulation is not from Alibaba. Do not budget on it.
No open-weight date or license. "Soon" has been the answer since July 19. No HuggingFace repository exists as of August 3. Kimi K3 shipped its weights on day 11. Qwen3.8 is on day 15 with no update.
What to Do Now
Test it yourself via Token Plan: The preview is accessible and reportedly strong on coding and agent tasks. Testing on your own workload is the only honest evaluation available — run it against the tasks that matter to your use case and compare to Kimi K3 and Grok 4.5 directly. The preview pricing (10% of standard) makes this cheap to evaluate.
Do not migrate production workloads yet: No standard API, no benchmark table, no model card, no license, no confirmed pricing. The five things worth watching before making any commitment: official benchmark table from Qwen, the active parameter count, a HuggingFace repository with a license file, published per-token API pricing, and independent evaluation from Artificial Analysis or LMArena.
How It Compares to What Is Available Now
| Model | API price /1M | Open weights | Independent benchmarks |
|---|
| Qwen3.8-Max (preview) | Not published | Promised — no date | None published |
| Kimi K3 | $3/$15/M | Live (July 27) | BenchLM: #5 of 214 |
| Grok 4.5 | $2/$6/M | Closed | Terminal-Bench 83.3% |
| DeepSeek V4 Flash | $0.14/$0.28/M | MIT (July 31) | Terminal-Bench 82.7% |
| Qwen3.7-Max | $2.50/$7.50/M | Closed | SWE-bench 80.4% |
Note: The $2/$6/M figure often cited for Qwen3.8-Max in X posts is the pricing for Grok 4.5 — not confirmed for Qwen3.8-Max. Qwen3.7-Max at $2.50/$7.50/M is the closest confirmed reference point.
Compare Qwen3.8-Max Against Other Models
Qwen Coding Prompts
While you are evaluating Qwen3.8-Max-Preview, these prompts work with both Qwen3.7-Max (production) and the preview: Best Qwen coding prompts 2026 →
Sources: MarkTechPost (July 19 announcement) · eesel AI hands-on review · Tech-Now.io analysis · emergent.sh independent review · BuildFastWithAI · Related: Kimi K3 — the open-weight model that shipped while Qwen3.8 is still preview → · DeepSeek V4 Flash 0731 — open weights, published benchmarks →