Qwen 3.8-Max Preview: 2.4T Params, Token Plan Pricing, and Open Weights Soon

Update — August 3, 2026: Qwen posted a launch video — “Meet Qwen3.8-Max: A New Bar for Coding and Cowork.” Open weights still not attached. Coding & cowork coverage →

Update — July 20, 2026: Moonshot paused new Kimi K3 subscriptions — overflow shoppers may route to Token Plan bundles (GLM + DeepSeek + Qwen). K3 subscription pause →

On July 19, 2026, Alibaba's official Qwen account posted that Qwen 3.8 is launching — a 2.4 trillion-parameter, continuously evolving stack that will go open-weight soon, while Qwen3.8-Max-Preview is already routable through the Token Plan, Qoder, and QoderWork.

The timing lands the same day Anthropic's Fable 5 promo extension expires — so developers comparing frontier coding subscriptions against China's subsidized model playbook have a fresh Alibaba option in the mix.

TL;DR — what people are asking

The official announcement (July 19, 2026)

Alibaba's Qwen team posted on X at 1:59 PM, July 19, 2026 (~119.5K views in the first hours):

Qwen3.8 is launching and going open-weight soon. 2.4T parameters, continuously evolving. One of the most powerful model available today, compatible to leading frontier AI models, second only to Fable 5. Qwen3.8-Max-Preview available NOW on Alibaba's Token Plan, Qoder, and QoderWork.

explainx.ai read: three distinct claims bundled together — scale (2.4T), access model (open-weight soon + preview API now), and frontier ranking (vs Fable 5). Only the third is easy to hype and hard to audit without public evals. The first two are actionable today: subscribe, route a harness, wait for weights.

Where to try preview (official pricing pages):

Open-weight soon vs Max-Preview now

Alibaba is running a two-speed launch familiar from GLM-5.2 and DeepSeek V4:

2.4T parameters signals this is not the next Qwen 3.6 27B local sweet spot. Even when weights ship, expect MoE routing, multi-node inference, or distilled smaller variants before laptop GGUF. Plan API-first; local-second.

"Continuously evolving" may mean rolling preview updates on the API — pin model IDs in production and re-run regression evals when Alibaba bumps the preview tag.

The Fable 5 comparison claim: what to believe

The post's boldest line — "second only to Fable 5" — deserves scrutiny, not dismissal.

Why skepticism is warranted:

1. No benchmark table in the announcement — unlike Qwen 3.7-Max's May launch, which cited Terminal Bench, SWE-bench Pro, and long-horizon demos.
2. Fable 5 access is policy-variable — export controls, ID verification, and usage cliffs mean many developers cannot A/B test against Fable even if they wanted to.
3. Vendor self-reporting inflates scores — enterprise buyers learned this from SWE-bench audit threads in July 2026.

Why the claim might still directionally hold:

• Qwen 3.7-Max already led Terminal Bench 2.0 (69.7) over DeepSeek V4 Pro (67.9) in May vendor tables.
• Alibaba invests heavily in environment scaling for agent generalization — 3.8 may be incremental on that stack.
• China's playbook rewards shipping capable models at subscription prices Western labs struggle to match.

Practical rule: if Qwen3.8-Max-Preview saves you money on your repo's agent tasks at acceptable quality, the ranking tweet is irrelevant. If it fails on multi-hour refactors, no slogan fixes that.

Token Plan pricing (Individual, July 2026)

Alibaba positions Token Plan as one subscription, multiple frontier models — not Qwen-only. Launch discounts apply on the international page as of July 19:

Models included on the plan (per pricing page): qwen3.7-max, Qwen3.8-Max-Preview (new), glm-5.2, deepseek-v4-pro, wan2.7-image-pro.

Team plans are listed as cheaper per seat than Individual — relevant for squads already comparing GLM Coding Plan at $9.99 or token economics across harnesses.

Credit mechanics: weekly pools plus 5-hour burst windows mean heavy agent sessions can hit burst limits before weekly caps — same class of friction as Claude usage limit resets. Start Lite for smoke tests; upgrade when burst pools become the bottleneck.

Supported coding tools and protocols

Token Plan exposes OpenAI-compatible and Anthropic-compatible endpoints — the same dual-protocol pattern that made Qwen 3.7-Max a drop-in for Claude Code and OpenClaw.

Officially listed supported tools:

Setup flow (all tools):

1. Subscribe on Token Plan (pick international or China URL).
2. Open API Keys — copy base URL and API key.
3. Paste into your harness provider settings.
4. Select Qwen3.8-Max-Preview (or qwen3.7-max for stable baseline) before long agent runs.

For GLM-style multi-harness examples, cross-read How to Run GLM 5.2 on every harness — endpoint wiring is parallel; only URLs and model IDs change.

Qwen Code setup (step by step)

Qwen Code is Alibaba's first-party agent shell — fastest path if you are not already on Claude Code or OpenCode.

1. Install Qwen Code from the official Qwen developer channel (same ecosystem as Qoder / QoderWork).
2. Open Settings → provider → Qwen Cloud Coding Plan.
3. Complete Token Plan checkout on your region's pricing page.
4. In the console, go to API Keys — copy base URL and key.
5. Paste credentials into Qwen Code provider fields; save.
6. Start a session and confirm the active model is Qwen3.8-Max-Preview.

Claude Code one-liner pattern (Anthropic protocol — adjust model ID to preview when listed):

export ANTHROPIC_BASE_URL="https://<your-token-plan-anthropic-endpoint>"
export ANTHROPIC_API_KEY="your-token-plan-key"
export ANTHROPIC_MODEL="qwen3.8-max-preview"
claude

OpenCode provider snippet (OpenAI-compatible — replace base URL from Keys page):

{
  "provider": {
    "qwen": {
      "name": "Qwen Token Plan",
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "https://<your-token-plan-openai-endpoint>/v1",
        "apiKey": "your-token-plan-key"
      },
      "models": {
        "qwen3.8-max-preview": { "name": "Qwen3.8-Max-Preview" }
      }
    }
  },
  "model": "qwen/qwen3.8-max-preview"
}

Verify routing: ask the model to state its ID, watch network calls hit Qwen endpoints (not api.anthropic.com), and run a small refactor before a multi-hour agent loop.

How Qwen 3.8 fits the stack you already have

China vs international: pick one region

Do not treat Token Plan as fungible across borders:

Reuters coverage of overseas AI restrictions and Anthropic–Alibaba distillation disputes are background noise for individual developers — but enterprise compliance teams should document which jurisdiction processes inference before routing proprietary code.

What to watch next

1. Open-weight release — watch Hugging Face Qwen org and top Chinese AI companies guide for GGUF publishers (Unsloth, bartowski) after official drops.
2. Independent benchmarks — Terminal Bench / SWE-bench entries for Qwen3.8-Max-Preview; compare to DeepSeek V4 Pro numbers.
3. Token Plan model roster — whether preview graduates to stable qwen3.8-max without price changes.
4. Distilled locals — if history repeats, a Qwen 3.8 30B-class dense variant may follow for local agent work months after the 2.4T flagship.

Related on explainx.ai

• Qwen3.8-Max coding & cowork video (Aug 3) — product pitch update; open weights still pending
• Kimi K3 subscription pause — Chinese AI capacity crunch (July 2026)
• Qwen 3.7-Max agent frontier — prior gen benchmarks and harness wiring
• Qwen 3.6 27B local dev — offline tier while waiting for 3.8 weights
• Fable 5 open-source alternatives — enterprise map when preview is not enough
• GLM 5.2 on every harness — parallel subscription setup patterns
• China AI playbook — why $6–68/mo subs exist
• DeepSeek V4 Pro coding — bundled on same Token Plan
• OpenCode local models guide — provider config reference
• Agent token economics — credit pools and burst limits

Official: Token Plan — international · Token Plan — China

Pricing, model roster, and preview availability reflect qwencloud.com and the Qwen X announcement as of July 19, 2026. Benchmark and Fable-ranking claims are vendor-stated until independently verified.