Frontline Hotspot
Frontline Hotspot

Alibaba's Qwen3.8-Max: 2.4T-param MoE flagship that programs autonomously for days

On 2026-08-03 Alibaba Tongyi released Qwen3.8-Max: a 2.4T-param MoE flagship with 1M context (991K input / 131K output), native vision across plan-execute-verify, positioned to "autonomously program for over ten days delivering complete projects." Pricing: ¥12/M input, ¥36/M output, explicit cache hit ¥1 (1/12 of uncached). Three entry points: blog / Qianwen platform / Qwen Studio.

Published August 3, 20264 min read
<!-- qwen3-8-max-hotspot | hotspot | Alibaba's Qwen3.8-Max: 2.4T-param MoE flagship that programs autonomously for days -->

On August 3, Alibaba's Tongyi officially released Qwen3.8-Max. This isn't a routine iteration-a 2.4-trillion-parameter MoE flagship, context pushed to 1M, and the official positioning is blunt: "can autonomously program for over ten days delivering complete projects." The model is no longer asked "can you?" but "are you done?"

Flagship specs: 2.4T-param MoE, 1M context

Qwen3.8-Max is a Mixture-of-Experts (MoE) architecture, 2.4 trillion total parameters, flagship tier. MoE means the model is large, but each inference activates only a subset of expert networks, so actual compute is far smaller than the parameter count-the key to scaling a model big without blowing up inference cost.

The context window is 1M tokens. A single turn takes up to 991K of input and returns 131K of output; in thinking mode input still reaches 983K with a max thought chain of 262K. Toss in a novel plus its reference docs and it reads while it works. Native vision understanding is standard: image, text, and video in, text out. The official line is that vision spans the full "plan-execute-verify" loop, supporting deep semantic parsing of ultra-long documents and long videos-not a glued-on OCR, but something built to work.

Pricing: cache hits drop to about a tenth

Pricing on the Qianwen AI platform:

ItemPrice (per million tokens)
Input¥12
Output¥36
Input (auto cache hit)¥1.5
Explicit cache creation¥15
Explicit cache hit¥1

The story is in the cache. Explicit cache hit at ¥1 versus uncached input at ¥12 is a 12x gap. Long-horizon tasks have high prefix repetition, so caching the shared prefix drives real input cost down to about a tenth. Explicit cache creation at ¥15 is 25% pricier than plain input, but you pay it once; subsequent hits run at ¥1. This is the cost switch built for "long conversation, multi-round iteration" scenarios-and Qwen3.8-Max is positioned exactly for long-horizon tasks, so the two fit tightly.

From "answering" to "working": long-horizon autonomous tasks

What's worth watching isn't the spec sheet, it's the shift in positioning. The official text reads: "coding and office capabilities comprehensively upgraded, can autonomously program for over ten days delivering complete projects," "handles legal, finance, design and hundreds of professional tasks, one conversation end-to-end delivering production-grade results."

Translated: the model is designed to take on a project that runs for over ten days-planning, executing, verifying, iterating on its own, not one round of Q&A. The supporting features are laid out for exactly this: function calling to wire external tools, structured output to guarantee JSON, web search for real-time info, batch tasks for scale, and fine-tuning open too. Partial Mode lets you give a prefix it strictly continues-good for pinning generation rhythm in engineering contexts.

"In long-horizon tasks, autonomous planning and closed-loop iteration, continuous evolution"-if that holds up under real load, it means the agent framework and the model itself are starting to ship fused together. Agents used to be a shell calling a model; now the model is growing its own agent bones.

How to try it right now

Three entry points, by use:

Take

The domestic flagship race is no longer about benchmark digits, it's about "how long a job can it finish." 2.4T params plus 1M context is the entry ticket; autonomous programming for over ten days with closed-loop iteration is the flag Qwen3.8-Max wants to plant. Pricing cache hits down to a tenth shows Alibaba has done the math: for long-horizon tasks to run at all, a single call has to be cheap enough to let it retry freely.

The real question is whether "ten-plus days to delivery" reproduces in real engineering. The positioning reads well, but a model planning a ten-day task chain on its own without drifting or collapsing mid-way is the hardest part-and that only shows when developers throw it into real codebases and run long-chain tasks. Today, claim an entry point and try it yourself; that beats any benchmark.


Sources

This article is AI-assisted and human-edited. Last updated: 2026-08-03

FAQ

How many parameters does Qwen3.8-Max have?
2.4 trillion parameters, MoE (Mixture-of-Experts) architecture, flagship tier. MoE means the total parameter count is large, but each inference activates only a subset of expert networks, so actual compute is far smaller than the parameter total-the key to scaling big without blowing up inference cost.
How long is the context window?
The context window is 1M tokens. A single turn takes max 991K input and 131K output; in thinking mode max input is 983K, max output 131K, with a max thought chain of 262K. Native vision supports image, text, and video input.
What is the API pricing?
Qianwen AI platform pricing: ¥12 per million tokens input, ¥36 per million output. Cache hits are far cheaper-auto cache hit ¥1.5, explicit cache hit ¥1 (versus ¥12 uncached input, a 12x gap). Explicit cache creation is ¥15, paid once with subsequent hits at ¥1, suited to prefix reuse in long-horizon tasks.

Related

Frontline Hotspot

280B Total, 14B Active, Three Regions Shut Out

On November 5, 2026 Baidu released ERNIE 5.5 Flash Base and published its weights: 280B total parameters with 14B activated per token, pretrained from scratch on 33.5T tokens, 128K training context and 1M for fine-tuning and inference, 43 layers with 37 heavily compressed attention layers plus 6 MLA-DSA layers, 512 routed experts with top-10 selection and one shared expert, and FP8 plus BF16 checkpoints on Hugging Face and ModelScope. This is an end-of-pretraining Base checkpoint with no mid- or post-training, not a chat model; a post-trained version is promised without a date. The custom ERNIE 5.5 Flash Base Model License Agreement v1.0 excludes the European Union, the United Kingdom and South Korea from its territory, allows free non-commercial use, requires a prominent Powered by ERNIE notice and a one-time notice for commercial use up to 50 million dollars in related annual revenue, requires separate prior written authorization above that threshold, and bans using outputs to improve unrelated AI systems.

Nov 6, 20269 min read
Frontline Hotspot

Kimi K3.1 Pro Weights Are Out: 2.8T Parameters, 1.56TB

At midnight Beijing time on November 5, 2026, Moonshot AI published the Kimi K3.1 Pro weights on Hugging Face, honoring the promise made when the model launched on October 23. The model has 2.8 trillion total parameters with about 104B activated per token, a 1048576-token context, 69 KDA plus 24 Gated DSA attention layers, a 401M vision encoder, and MXFP4 weights with MXFP8 activations under quantization-aware training, totaling roughly 1.56TB. The vendor-reported Terminal-Bench 4.0 score is 65.7; on the third-party Agent Arena of real user sessions it ranks first among open models with a net improvement rate of 5.29 percent, ahead of MiMo-V2.6-Pro at 2.68 percent (preliminary). The Kimi K3.1 Pro License requires a separate agreement for Model-as-a-Service businesses exceeding 20 million dollars in trailing 12-month revenue and prominent naming for products above 100 million monthly users or 20 million dollars monthly revenue.

Nov 5, 20269 min read
Frontline Hotspot

iFlytek Spark X2.6: A 293B Flagship, but Only a 4B Is Open

iFlytek launched Spark X2.6 at its global 1024 Developer Festival in Beijing on October 22, 2026: a 293B-parameter MoE that activates 30B per token, a 1M-token context window, an end-cloud integrated design, trained entirely on domestic compute and served from a 10,000-card domestic inference cluster. The efficiency ledger, all vendor-reported: training 2000B tokens of 256K-length data on the Feixing-2 platform fell from 4.5 months to 1.2 months (a 4x gain), and the first fully domestic four-layer KV Cache architecture lifted hit rate from 63 to 95.7 percent with 4.5x inference efficiency. Self-reported scores put SWE Pro at 77.3 and SWE Multilingual at 95.0, both first, with SWE Verified at 95.4 behind only DeepSeek-V4.1-Flash 1017's 95.6 - yet there is no complete score table, no technical report, and no third-party replication. The key correction: the viral headline 'iFlytek open-sources X2.6' misreads the release. Only the on-device Spark-X2.6-4B is open under Apache 2.0 (Hugging Face under the XHToken namespace, ModelScope mirror, Ollama as SparkLLM), while the 293B flagship stays API-only. The 4B self-reports six gains including BFCL-V4 71.8 and SWE Verified 62.8. Launch pricing is 0.9 / 0.08 / 3.6 yuan per million tokens. The narrative layer predicts 60 percent local, 30 percent hybrid, 10 percent cloud workloads with routing accuracy up from 81.2 to 96.1 percent; the chairperson's 'only model trained on fully domestic compute' is a unilateral claim with no chip model disclosed. Unpublished: chip models, cluster size, total FLOPs, technical report, flagship weights.

Oct 23, 20268 min read