Qwen 3.8 Preview — 2.4T Multimodal, and What's *Not* Yet Public
A rundown of Qwen 3.8 (Max Preview), Alibaba's 2.4-trillion-parameter multimodal model announced days after Kimi K3. What's confirmed (2.4T sparse MoE, ~1M context, text/image/video/document multimodal) versus what's still undisclosed (active parameters, license, model card, official benchmarks, pricing) — and how to read the announcement.
Days after Kimi K3 shook the market as 2.8T open weights, Alibaba fired right back with Qwen 3.8. Announced on X on July 19, 2026, it's a 2.4-trillion-parameter (2.4T) multimodal model that Alibaba claims sits "just below Claude Fable 5" on internal evals. But this is a preview, not a full release — open weights are promised "soon," with no date, license, model card, official benchmarks, or pricing yet. So this post focuses less on spec-flexing and more on soberly separating what the announcement confirmed from what's still blank.
What you'll learn
- The confirmed specs in the Qwen 3.8 preview versus the still-undisclosed items
- How it contrasts with Kimi K3 in the same "open trillion-scale MoE" wave
- How to (cautiously) read the announced benchmark claims
- Why the undisclosed active-parameter count makes self-hosting decisions hard
- What to do now, and what to wait for
⚠️ The following reflects the July 2026 preview announcement and early reporting. Many values are unfinalized at preview stage, and the benchmarks are mostly self-claims or small community runs. Confirm the official model card and license before making adoption decisions. For open models broadly, see the 2026 AI model landscape.
1. What the Qwen 3.8 Preview Confirmed
First, the specs actually disclosed in the announcement.
- Parameters: 2.4 trillion (2.4T), sparse Mixture-of-Experts (MoE)
- Multimodal: native across text, images, video, and documents
- Context: ~1M-token class (~983,616 tokens per Qwen Cloud metadata), max output ~131,072 (128K) tokens
- Variant: Qwen3.8-Max (currently
Qwen3.8-Max-Preview) - Improvement claims: gains over the prior generation in coding, full-stack development, data analysis, and office workflows
- Access: preview available on Token Plan, Qoder, and QoderWork
- Announced: 2026-07-19 (X); open weights teased "soon"
2. What's Not Yet Public
This is the crux of the post. Being a preview announcement, many cells are empty — and these blanks directly affect adoption decisions.
| Item | Status | Why it matters |
|---|---|---|
| Active parameters | ❌ undisclosed | In sparse MoE, the single most important number for per-token compute and serving cost |
| Open-weight release date | ❌ unset ("soon") | It says "open weights," but you can't download it yet |
| License | ❌ undisclosed | Whether commercial use/redistribution is allowed is undecided |
| Model card | ❌ undisclosed | No official doc on training, evaluation, safety, or limits |
| Official benchmark table | ❌ undisclosed | No basis to verify the performance claims |
| Standard per-token pricing | ❌ undisclosed | TCO can't be computed |
In one sentence: "announced ≠ usable." Qwen 3.8 is currently a preview, and "open weights" is a promise, not a shipped artifact.
3. Contrast With Kimi K3
The two announcements, only days apart, are different in nature. Positioned within the same "open trillion-scale MoE" wave:
| Item | Qwen 3.8 (Max Preview) | Kimi K3 |
|---|---|---|
| Parameters (total) | 2.4T (sparse MoE) | 2.8T (MoE) |
| Active parameters | undisclosed | 16 / 896 experts (~1.8%) |
| Context | ~984K (1M-class) | 1M |
| Multimodal | text, image, video, document | text, image (visual) |
| Open weights | "soon" (no date) | release date stated |
| Benchmarks | no official table (self-claims) | independent eval numbers exist |
| Status | preview | full announcement |
The key difference is verifiability. Kimi K3 has independent eval numbers and a stated weight-release date, whereas Qwen 3.8 is, for now, mostly Alibaba's own claims. Both call themselves "open weights," but the moment you can actually hold them — and the thickness of the evidence — differ.
4. Claimed Strengths — Where It Says It Shines
Alibaba's highlighted improvements cluster around development, data, and work automation.
- Coding & full-stack development — claimed gains over the prior generation, surfaced alongside agentic dev tools (Qoder, etc.).
- Data analysis — multimodal understanding of tables/documents maps onto data-pipeline contexts.
- Office workflows — document processing and task automation.
- Video & document multimodal — claiming a broader input modality than Kimi K3 (mostly image) is a differentiator.
That said, all of this is currently self-claimed and unconfirmed by independent benchmarks.
5. How to Read the Benchmarks
Be especially careful with the performance story.
- Alibaba's internal eval: claims a level "just below Fable 5."
- Community runs (e.g., KingBench 3): some reports put Qwen3.8-Max above Opus 4.8 and Kimi K3, with only Fable 5 ahead — but these are small user-run samples, not official tables.
- Independent third-party benchmarks: currently none.
In other words, the rankings floating around are self-claims or small-sample runs. Treat preview-stage performance numbers as "signals of potential" only; it's too early to base an adoption decision on them.
⚠️ Benchmark leaderboards have limits of their own. A high score doesn't mean effective performance on your workload. Evaluating on your own tasks is always best.
6. Self-Hosting View — Why the Math Doesn't Close Yet
Judging by the 2.4T total alone, you'd expect the same serving burden covered in the Kimi K3 post. But a decisive variable is missing.
With active parameters undisclosed, you can't estimate per-token compute (FLOPs) or serving cost.
In sparse MoE, total parameters drive the memory (VRAM) burden while active parameters drive the compute burden. Kimi K3 disclosed "16/896 active," enabling the "memory is large but per-token compute is light" calculation — but for Qwen 3.8 half of that equation is blank. On top of that, with no license or weights yet, you can't even plan self-hosting to begin with. Use the weight-VRAM approximation (FP16 ~2GB per 1B, 4-bit ~0.5GB) to get a rough sense of the total, but defer any concrete serving design until the active-parameter count and quantization format are public.
7. Wrap-up — What to Do Now, What to Wait For
Qwen 3.8 is significant as a signal that the "open trillion-scale MoE" race is heating up week by week. But given its preview status, the right moves now are clear.
| Situation | Recommendation now |
|---|---|
| Adopt into production immediately | Too early — wait for the full release, license, and benchmarks |
| Gauge capability early | Evaluate on your own tasks via the preview (Token Plan/Qoder) |
| Consider open-weight self-hosting | Hold until active parameters, license, and weights are public |
| Need stable open weights now | Consider options already shipped, like Kimi K3 |
Bottom line: Qwen 3.8 is at the "watch, but don't bet yet" stage. Use the preview to get a feel; judge once the full release brings a license, model card, and independent benchmarks.
For the broader context, continue with the 2026 AI model landscape and the open-source LLM comparison. If you're evaluating an on-prem LLM rollout, get in touch.
Glossary
The terms used in this post, in one place. For broader AI terminology, see the AI glossary.
| Term | Meaning |
|---|---|
| Frontier Model | The most capable class of models at a given moment (e.g., Claude Fable 5/Opus, GPT). Qwen 3.8 claims to sit "just below" the frontier. |
| Open-Weight (Korean: 오픈웨이트/오픈메이트) | A model whose trained weights are published so you can download and run/serve it yourself. But when only a "coming soon" promise exists (as with Qwen 3.8), you can't actually hold it yet. |
| Preview | A pre-release stage shipped to some channels first, incomplete — license, model card, and final benchmarks may not exist yet. |
| MoE (Mixture-of-Experts) | An architecture splitting the model into expert sub-networks and firing only a few per token. Total parameters are large, but per-token compute uses only the active ones. |
| Sparse MoE | MoE that activates only a small fraction of all experts. Qwen 3.8 uses this, but the active fraction is undisclosed. |
| Active Parameters | The parameters that actually participate in computing a given token. The basis for compute (FLOPs) and serving cost. Qwen 3.8 has not disclosed this. |
| Multimodal | The ability to take in and understand multiple formats — text plus images, video, documents. Qwen 3.8 claims native multimodality. |
| Context Window | The maximum token length a model can attend to at once. Qwen 3.8 is ~984K (1M-class). |
| Max Output | The maximum tokens generatable in a single response. Qwen 3.8 is ~128K. |
| Model Card | The official document covering training data, evaluation, safety, limits, and license. May not exist at preview stage. |
| License | The legal terms defining permitted commercial use, redistribution, and modification. If undisclosed, real-world usability is uncertain. |
| Benchmark | A score measuring performance on standard tasks. Self-claims, community runs, and independent third-party checks differ in trustworthiness — and a high score isn't the same as performance on your workload. |
| Quantization | Lowering weight precision (FP16 → 8/4-bit) to reduce memory and compute. The key means of fitting large models into realistic VRAM. |
| VRAM | GPU memory. It must hold weights + KV cache + activations, and is the primary bottleneck for serving trillion-scale models. |
| FLOPs | Floating-point operations — an approximate proxy for inference cost/speed; in MoE, measured against active parameters. |
| TCO (Total Cost of Ownership) | The total cost to adopt and operate (GPUs, power, ops staff, etc.). Can't be computed without pricing and active-parameter data. |