Qwen3.8-Max's Launch Plan: API First, Weights Separately
MODELS & COSTS
Qwen3.8-Max's Launch Plan: API First, Weights Separately
The API release and the promise of Max-class weights were different milestones. A self-hosting plan needed the actual checkpoint and final license, not just the announced parameter count. This launch account keeps that missing artifact visible beside Qwen's published demonstrations.
In this article
Earlier mid-size Qwen releases offered Apache 2.0 weights while Max remained API-only. A downloadable Max checkpoint would expand self-hosting options, but its license and hardware requirements determine who can use it. The announcement alone does not establish those terms.
▲ Performance summary chart from the official Qwen blog (source: qwen.ai/blog?id=qwen3.8)
Timeline: two weeks from preview to release
Here's the summary. On July 19, during Shanghai's WAIC, Qwen3.8-Max-Preview made a surprise appearance — conveniently two days after Moonshot dropped Kimi K3 (2.8T total parameters, open-weight). Back then there wasn't a single benchmark table, just the claim that it was "the strongest thing after Fable 5," so skepticism was high. Then two weeks later, on August 2, the official version arrived with the full benchmark table and showcases. Even Bloomberg reported that it "beats Kimi K3 on some benchmarks."
| Item | Details |
|---|---|
| Model | Qwen3.8-Max (based on the Qwen 3.5 architecture) |
| Parameters | 2.4T total / 95B active per token (sparse MoE) |
| Modality | Native multimodal (text·image·video·document) — first multimodal model above 1T |
| Context | 1,000,000 token input / up to 65,536 token output |
| Price | Input $2.0 / Output $6.0 / Cached input $0.25 (per 1M tokens) |
| Open weights | Coming to Hugging Face + ModelScope next week (Qwen3.8-27B alongside) |
| Access | QwenCloud API (OpenAI & Anthropic protocol compatible), Qwen Studio |
Qwen announced 2.4T total parameters and 95B active per token. The total belongs in the storage calculation; the active count describes part of the per-token computation. At this launch stage, those specifications did not establish an available weight format or a complete local memory requirement.
▲ Total-parameter comparison. 2.4T is 1.5× the previous largest, DeepSeek V4 Pro (1.6T)
The vendor's coding-agent evaluations
From the full table on the official blog, I pulled just the coding-agent entries. The comparison set is Claude Opus 4.8, Fable 5, GPT-5.6 Sol, and the previous generation, Qwen3.7-Max.
▲ Vendor-reported numbers. Fable 5 shown as reference.
The generational jump is obvious. DeepSWE 1.1 jumped 21.6 → 56.6, FrontierSWE 40.7 → 73.5. PaperBench hit 93.0, passing Fable 5 (88.8), and Terminal Bench 2.1 at 86.6 is front-of-pack too. On the multimodal side, some entries also lead the frontier models — BabyVision 82.0, OSWorld-Verified 86.1.
But there's stuff to filter out. A good chunk of the table is Qwen's own in-house benchmarks — QwenSWEBench, QwenQoderBench, QwenBlenderBench, things like that. And most measurements are vendor self-runs on a Claude Code harness. "Ranked #2 after Fable 5" is vendor positioning, not independent verification. We'll see third-party scores from Artificial Analysis or LMArena when they land.
Long-running task demonstrations
The real play here wasn't the numbers — it was the showcases. Three cases, and one of them has its full commit history public on GitHub so you can check it yourself.
| Case | What happened | Result |
|---|---|---|
| Autonomous coding | Ran the oh-my-cli project fully autonomously for 16 days from an empty repo (issue intake → implementation → tests → merge loop) | 265 commits, 127 PRs, 151 issues — repo is public |
| Paper reproduction + improvement | Given a single data-selection paper, told to reproduce then improve. ~125 hours, 7,600 lines of code, 33 GPU training rounds | Reproduced the method, then ran 18 experiments with its own ideas → AIME24 +2.71p |
| Live competition | Tianchi multimodal intent-recognition challenge, same conditions as 526 human teams, 24-hour limit | Accuracy 0.60 → 0.853, beat 458 teams (87%) |
Personally, the chip-design case was the most fun. It was told to design a GCD/RSA crypto hardware accelerator from scratch, and over 500 autonomous turns it cut the gate count from 8,298 to 678 — a 92% reduction. Physical layout also shrank from 106×106µm² to 46×46µm², and it hit 500MHz timing closure. What impressed me was an algorithm-level rewrite mid-run: "replace the modulo division hardware with a shift-subtract iterative structure."
▲ Gate count over autonomous turns (reconstructed from Qwen's official announcement data)
For long-horizon operation, they ran a 365-day e-commerce simulation benchmark on real Taobao/Tmall transaction data: starting capital ¥100,000 grew to ¥416,252 (4.16×). That's 38% higher than #2 GLM-5.2, in an environment that includes 152 fake-supplier traps, typhoons, and stockouts. Of course — vendor benchmark again, take it as such.
Launch API pricing
| Model | Input ($/1M) | Output ($/1M) | Cached input ($/1M) |
|---|---|---|---|
| Qwen3.7-Max | $2.50 | $7.50 | 90% discount |
| Qwen3.8-Max | $2.00 | $6.00 | $0.25 |
The generation went up and token prices went down 20%. That's the 95B-active sparse MoE structure — serving costs are just lower. The API is compatible with both OpenAI and Anthropic protocols on QwenCloud, so it plugs straight into existing agent harnesses like Claude Code, Codex, OpenClaw. Fun detail: Hermes has its own spot on the harness-compatibility chart in the official blog — which is the environment I'm writing this in, so that was a nice little moment.
Checkpoint, license and capacity requirements
The announcement gives a 'next week' target, also reported by Bloomberg. That is a planned release date rather than evidence of a published artifact. Three conditions still matter:
First, the license isn't out yet. The 3.5/3.6 precedent is Apache 2.0, but slapping a restrictive community license on a 2.4T model would still technically count as "open-weight." We need to see the actual license file.
At hypothetical uniform 4-bit precision, 2.4T parameters alone occupy about 1.2TB before overhead. That is beyond a typical single consumer system, although distributed and offloaded configurations need their own tests. Open weights could support provider competition and controlled self-hosting; lower token prices are a possible outcome, not a guaranteed consequence.
Third, every number is vendor-reported. Until independent verification shows up, treat it as direction, not fact.
Before planning self-hosting, confirm that the exact checkpoint is downloadable and read its license. For API use, verify the active model version and current rate in the console. The launch plan should not be mistaken for delivery of the weights.
Sources: Qwen official blog (qwen.ai/blog?id=qwen3.8, 2026-08-02) · @Alibaba_Qwen announcement · MarkTechPost · techsy.io · Bloomberg (2026-08-03). All benchmark figures are vendor-reported.
Related on this blog: Qwen3.8-Flash-Next: 6B Active, Qwen4's Architecture Early · Qwen3.8-27B: 27B Dense, Vision Included, Apache-2.0
Comments
Post a Comment