$2/M input tokens, $6/M output tokens.Official Release: August 3, 2026
Three weeks after its initial preview announcement at the World Artificial Intelligence Conference in Shanghai, Alibaba Cloud officially released Qwen3.8-Max on August 3, 2026, making it available through the QwenCloud API platform under the model ID qwen3.8-max. The same release announcement confirmed that Qwen3.8-27B — a smaller second checkpoint — would ship as open weights alongside the Max variant.
Qwen3.8-Max is built upon the architectural foundation of Qwen 3.5 and represents Qwen's most capable model to date. It is a Mixture-of-Experts (MoE) model with 2.4 trillion total parameters, supporting text, image, and video input with text output. Hybrid thinking mode is enabled by default.
The QwenCloud API endpoint uses the model ID qwen3.8-max. Third-party clients compatible with OpenAI or Anthropic protocols can connect directly once an API key is issued through the QwenCloud developer console.
Qwen3.8-Max: Confirmed Specifications
Here is every specification that has been confirmed through official QwenCloud documentation and the August 3 release announcement:
| Specification | Value | Status |
|---|---|---|
| Total parameters | 2.4 trillion | ✅ Confirmed by Alibaba |
| Architecture | Sparse MoE | ✅ Confirmed |
| Activated parameters per token | 95 billion | ✅ Confirmed (research sources) |
| Context window | ~983,616 tokens (~1M) | ✅ Confirmed (QwenCloud docs) |
| Max output length | 131,072 tokens | ✅ Confirmed (QwenCloud docs) |
| Modalities | Text + Image + Video → Text | ✅ Confirmed |
| Thinking mode | Hybrid (enabled by default) | ✅ Confirmed (QwenCloud docs) |
| Number of experts | Not disclosed | ❌ Unknown |
| Experts activated per token | Not disclosed | ❌ Unknown |
| Training data | Not disclosed | ❌ Unknown |
| Published benchmarks | None | ❌ None |
| Base architecture | Qwen 3.5 | ✅ Confirmed (OpenLM.ai) |
Qwen3.8-Max is described by Alibaba as delivering "comprehensive improvements across coding, work, research, and long-horizon tasks" over the Qwen 3.7 series. It is positioned as the most capable model in the Qwen family to date.
Qwen3.8-27B: The Smaller Variant
Alongside Qwen3.8-Max, Alibaba confirmed a second checkpoint called Qwen3.8-27B. This model is significantly smaller — 27 billion parameters — and was explicitly announced as going open-weights.
As of this writing, the following is known about Qwen3.8-27B:
- Parameter count: 27 billion (confirmed by name)
- Architecture type: Unknown — could be dense or MoE. The 27B size is unusual; Qwen3's dense lineup includes 14B and 32B, but no 27B. This suggests it may be a specialized variant or a renamed checkpoint.
- Open-weight status: Promised, but no HuggingFace repository, model card, or license file has been published as of August 3, 2026.
- Benchmarks: None published.
- Context window: Unknown.
- Relationship to Qwen3.6-27B: Unclear. Qwen3.6-27B (released April 22, 2026) is a dense 27B model with vision capability that scored 77.2% on SWE-bench Verified, beating Qwen3.5-397B-A17B on that benchmark. Qwen3.8-27B may be an updated version built on the Qwen3.8 architectural foundation.
The most important unanswered question about Qwen3.8-27B is whether it is dense or MoE. A 27B dense model would be a significant step down from Qwen3.8-Max's 2.4T, but would be immediately runnable on consumer hardware. A 27B MoE model with a much larger total parameter count and only a fraction activated per token would be a completely different product. Until the model card is published, the 27B name alone is ambiguous.
Pricing: API Costs
Qwen3.8-Max pricing through QwenCloud is:
| Input Type | Price per 1M Tokens |
|---|---|
| Cached input | $0.25 |
| Standard input | $2.00 |
| Output | $6.00 |
QwenCloud's pricing page notes that billing varies by model type: text models charge per token, image generation per image, and video generation per second. Failed API calls are not charged.
Qwen3.8-27B pricing has not been published. If it follows the pattern of smaller Qwen models, it would likely be significantly cheaper than the Max variant, but no official figures exist yet.
Open-Weight Status: Promised, Not Delivered
The open-weight promise is the most consequential claim about Qwen3.8. Here's the current status:
| Artifact | Status | Details |
|---|---|---|
| Qwen3.8-Max weights | Promised | No HF repo, no model card, no license. Promised "next week" from August 3. |
| Qwen3.8-27B weights | Promised | No HF repo, no model card, no license. Confirmed as open-weights target. |
| Model cards | Not published | No technical specifications beyond parameter count and architecture type. |
| License | Not published | Unknown — Apache 2.0, Qwen license, or proprietary? |
| Quantized formats (GGUF, 4-bit) | Not available | No Unsloth or community uploads yet. |
Alibaba has a pattern of announcing open-weight releases for Max-tier models that don't materialize. Qwen3.7-Max and Qwen3.6-Max-Preview were both announced with open-weight promises but remained proprietary API-only models. The Qwen3.8 open-weight promise is the first time a Max-class model has been explicitly paired with a smaller open-weight checkpoint (Qwen3.8-27B), which is a meaningful shift in strategy — but patterns matter more than promises.
Unsloth Support: Not Yet Available
Unsloth, the popular fine-tuning and inference optimization platform, has not yet uploaded any Qwen3.8 models to HuggingFace.
Unsloth's current Qwen3 collection includes models from the original Qwen3 (2025) release and Qwen3.6, available in GGUF, 4-bit, and 16-bit Safetensor formats using Unsloth Dynamic 2.0. Unsloth's GitHub README currently lists support for: Kimi K3, Gemma 4, Qwen3.6, DeepSeek-V4, and GLM — with no mention of Qwen3.8.
Once Qwen3.8 open weights are released, Unsloth is expected to add support following their standard pattern: 16-bit and 4-bit quantized uploads, GGUF variants, and fine-tuning support with their Dynamic 2.0 methodology. This typically happens within days of the official release, based on their track record with previous Qwen models.
When Qwen3.8 weights go live, check huggingface.co/collections/unsloth/qwen3 for new uploads. Look for 4-bit bnb formats for fine-tuning on consumer GPUs and GGUF variants for local inference with llama.cpp or Ollama.
What's Missing: The Gaps to Fill
Before Qwen3.8 can be properly evaluated or deployed, the following information needs to be published:
- Model card and technical report — Training data description, evaluation methodology, limitations, and known failure modes.
- Published benchmarks — GPQA, SWE-bench Verified, HumanEval, and other standard evaluation suites.
- License file — Determines what you can legally do with the weights once they're released.
- MoE architecture details — Total number of experts, experts activated per token, expert routing strategy.
- Qwen3.8-27B full specifications — Architecture type (dense vs MoE), context window, training data, and benchmarks.
Until these items are published, Qwen3.8 remains a model with impressive claimed specifications and no verifiable evidence of performance.
Qwen3.8 vs Qwen3.6: Where We Stand
Qwen3.6-27B (released April 22, 2026) is the most recent open-weight general model available today. Here's how the two compare with confirmed data:
| Specification | Qwen3.6-27B | Qwen3.8-27B |
|---|---|---|
| Parameters | 27B | 27B |
| Architecture | Dense | Unknown |
| Vision | Yes | Unknown |
| SWE-bench Verified | 77.2% | Not published |
| Open weights | Yes | Promised |
| Available now | Yes | No |
Qwen3.6-27B's 77.2% on SWE-bench Verified is notable because it beats Qwen3.5-397B-A17B (the previous Qwen flagship) on that benchmark while using a fraction of the hardware. If Qwen3.8-27B improves on that while maintaining similar hardware requirements, it would be a significant achievement.