基本信息

要点解读

这是什么

这是一条关于阿里巴巴 Qwen 发布新旗舰模型 Qwen 3.8 Max 及其 27B 版本的消息，强调超大规模参数、长期自主编码和多模态代理能力，提供付费 API 并承诺随后开放权重。

用在哪里

适用于需要强大开源模型进行代码生成、硬件设计仿真、企业级业务流程自动化以及多模态交互的研发团队和研究人员，也适合关注大模型在自主代理和视觉反馈表现的评测者。

可以推断的

推测：该模型的高参数规模和自主工作能力可能吸引需要大规模本地部署或定制化微调的用户，尤其是关注长期任务自动化的开发者。
推测：由于提供 API 并承诺开放权重，模型有望在开源社区快速产生第三方工具和插件生态。

来源摘要/节选

After the Qwen Exodus last year and new management took over launching more closed model APIs, there was some real doubt as to whether or not this leading open models lab would continue to release relevant models.

That doubt is now gone. Qwen 3.8 Max is a MONSTER 2.4T model that would have been the top open model in the world but for the Kimi K3 release we already covered.

Qwen offers them on API for $2 input/$6 output per million tokens, but they have promised to open-weight both models.

Key Capabilities & Breakthrough Highlights

Autonomous Long-Horizon Coding:

10+ Days Unattended Coding: Built a self-evolving coding harness from scratch over a multi-week autonomous run.

Autonomous AI Research: Rebuilt a complete paper’s pipeline (Unified Data Selection for LLM Reasoning) from scratch, then autonomously ran an iterative research loop over 125 hours to invent a new data selection method beating the original paper’s benchmark by +2.71 points.

Competitive Data Science: Competed against 526 human teams in the WWW2025 Multimodal Dialogue Intent Recognition Challenge, placing in the top 13% (outperforming 87% of human teams) within 24 hours.

Autonomous Hardware & Chip Design:

Executed a complete silicon design flow (GCD/RSA cryptographic accelerator) from RTL editing to simulation, synthesis, and physical layout.

Reduced gate count from 8,298 to 678 gates while achieving an 81% die area reduction and meeting physical timing closure at 500 MHz.

Deep Real-World Work & Operations:

Demonstrated production-grade outputs across hundreds of professional workflows (e.g., corporate legal reviews, UI/UX design, structural engineering models, and automated ETF quant research).

Outperformed competing models in the E-Commerce Bench (a 365-day store operation simulation), generating a 4.16x return (¥416,252 balance) through continuous game-theoretic negotiation and inventory planning.

Multimodal Agents & Visual Feedback:

Integrates native visual feedback across planning, coding, and GUI interaction, enabling direct application recreation across platforms (desktop, mobile, web).

Released Qwen-MM-Plugins to extend multimodal capabilities to existing agent frameworks.

A very nice win for open weights! On today’s pod with Baseten we talked about what it’s like to support these massive model drops on release.

AI News for 7/25/2026-7/27/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Top Story: Qwen 3.8 Max open model launch

What happened

Alibaba Qwen announced Qwen3.8-Max as its new flagship and said open weights are coming next week.

Alibaba introduced Qwen3.8-Max as its “most capable model to date,” describing it as a 2.4T-parameter model focused on coding, long-horizon agentic work, and multimodal reasoning, with the explicit claim that open weights will be released next week, alongside Qwen3.8-27B also going open-weight @Alibaba_Qwen

The launch tweet also included API pricing: $2.00 / M input tokens, $6.00 / M output tokens, and $0.25 / M cached tokens @Alibaba_Qwen

Alibaba framed the model around several headline capabilities: 10+ days of autonomous coding, 500+ turns of chip design optimization, 365 days of e-commerce strategy, and native multimodal intelligence where vision is part of the execution loop rather than just an input channel @Alibaba_Qwen

The company simultaneously pushed availability across its own surfaces and partners: Qwen Studio, API, Command Code, and later Venice; infra and app builders quickly confirmed support plans or integrations including Baseten, Hermes Agent, and Command Code @Alibaba_Qwen @Alibaba_Qwen @baseten @Teknium

The announcement landed as part of a broader pattern: multiple observers described it as evidence that the Chinese open-weight frontier is now competing directly with top Western closed models, especially in coding, agentic workflows, and multimodal tasks @kimmonismus @matvelloso

Official claims and reported specs

Vendor-reported model details and performance claims were unusually aggressive for an open-weight release.

Alibaba’s own framing:

2.4T total parameters @Alibaba_Qwen

Long-horizon agentic/cowork focus @Alibaba_Qwen

Autonomous coding over 10+ days with a public GitHub trace @Alibaba_Qwen

500+ turns for chip design optimization @Alibaba_Qwen

365 days of e-commerce strategy execution @Alibaba_Qwen

Native multimodal planning loop rather than vision-only input @Alibaba_Qwen

Third-party summary tweet from ZhihuFrontier added more claimed or reported technical details:

95B active parameters per token, implying an MoE activation ratio of roughly 4%

1M-token context window

API exposes low / medium / xhigh reasoning-effort modes

Compatibility with OpenAI and Anthropic protocols

Benchmark claims: PaperBench 93.0, CoWorkBench 74.8, WideSearch 81.9 @ZhihuFrontier

Vals AI independently posted concrete eval/runtime settings:

1M token context

128k max output tokens

Tested at temperature 0.7 with default top-p / top-k @ValsAI

These numbers matter because they place Qwen3.8-Max in the same deployment class as other giant sparse open models like Kimi K3 and GLM-5.2, not the more practical 30B–70B local tier.

Independent evaluations and leaderboard placements

The model immediately posted strong third-party results, especially in coding-adjacent, vision, and design-heavy arenas.

Frontend Code Arena: Qwen3.8-Max debuted at #4 overall with 1,668 Elo, trailing only Claude Opus 5 [Max] at 1,705 and Kimi K3 [Max] at 1,676, and roughly tied with Claude Opus 5 [High] at 1,669 @arena

In Frontend Code Arena subslices, it ranked:

#2 Consumer Product

#3 Brand & Marketing, Reference-based design, Gaming, Content Creation Tools

#4 Data & Analytics

#5 Simulations @arena

Vision Arena: Qwen3.8-Max ranked #2 with 1,305, only 13 points behind Claude Fable 5 [High] @arena

Vals Index: Qwen3.8-Max ranked #2 among open-weight models, #10 overall out of 43, with a score of 66.1 @ValsAI

Vals also reported:

It matched Claude Opus 4.7 on the Index, 66.1 vs 66.1

At about 2.3x lower cost per test: $2.68 vs $6.17 @ValsAI

Vals’ benchmark-specific numbers:

SWE-bench: 87.3%, ahead of GPT-5.5 (82.6%) and GLM-5.2 (83.3%), but behind Claude Opus 4.8 (89.2%)

Terminal-Bench 2.1: 67.4, up from 61.0 for Qwen 3.7 Max @ValsAI

Vals also highlighted the pace of progress:

Qwen 3.7 Max = 57.5

Qwen 3.8 Max = 66.1

Gain of 8.6 points in ~2.5 months

Price cut from $2.50/$7.50 to $2.00/$6.00 input/output @ValsAI

There were also more anecdotal but technically relevant claims:

One user visualized benchmark deltas and argued “Opus 4.8 is mostly subsumed by 3.8-Max” on the chart they reconstructed @deliprao

Another claimed Qwen 3.8 surpassed Fable 5 on Terminal Bench and said Anthropic was now under visible pressure @kimmonismus

A separate tweet called Qwen 3.8 Max the “best object detection VLM” across satellite, infrared, documents, technical drawings, sketches, crowded scenes, and small objects, though this was based on examples rather than a cited benchmark paper @skalskip92

Facts vs. opinions

Facts / directly attributable claims

Alibaba announced Qwen3.8-Max and said open weights arrive next week; Qwen3.8-27B will also go open-weight @Alibaba_Qwen

Alibaba disclosed API pricing of $2 input / $6 output / $0.25 cached per million tokens @Alibaba_Qwen

Arena reported #4 in Frontend Code Arena at 1,668 and #2 in Vision Arena at 1,305 @arena @arena

Vals reported 66.1 on Vals Index, #2 among open-weight models, 87.3% SWE-bench, 67.4 Terminal-Bench 2.1, 1M context, 128k output, and lower cost-per-test than Opus 4.7 @ValsAI @ValsAI @ValsAI

ZhihuFrontier stated 95B active parameters and protocol compatibility; this appears to be a secondary summary rather than an original Alibaba spec sheet @ZhihuFrontier

Opinions / extrapolations / rhetoric

“China is no longer lagging behind but competing on equal footing” @kimmonismus

“Open models are winning now” @JonathanRoss321

“Looks like Opus 4.8 is mostly subsumed” @deliprao

“Anthropic is under pressure” and “mood shifted drastically” are ecosystem readings, not measurements @kimmonismus

“Best object detection VLM” is an informed product judgment, but not one tied in-thread to a standard benchmark table @skalskip92

Claims that Qwen3.8-Max plus open agents prove open models have “caught up” are user-level interpretations rather than consensus eval conclusions @omarsar0

The central factual story is strong even after stripping out the hype: a very large sparse model, open-weight promise, lower pricing than prior Qwen Max, and high placements on multiple third-party leaderboards.

The infrastructure reality: “open-weight” does not mean easy to run

A major counterpoint in the discussion was that frontier open models are operationally open, but not broadly accessible in the local-inference sense.

Jamin Ball argued that pricing comparisons were overstated because “vanilla” token prices ignore token efficiency and because these models are enormous:

Qwen 3.8 Max >2T params

Kimi K3 ~104B active per token

GLM 5.2 = 744B total, 40B active

For K3, loading weights alone is >1TB memory

Requires at least 8 H100/B200 GPUs to run

Moonshot recommends 64+ accelerators in supernode-style setups @jaminball

This same critique implicitly applies to Qwen3.8-Max, even if its active-parameter count is somewhat lower than K3’s: a 2.4T-class MoE is not a commodity local model @jaminball

StableQuan made the practical version of the same point more bluntly: long, RAM-heavy prompts and slow tool calls make giant models painful on consumer hardware, recommending API use instead @stablequan

At the same time, the excitement around Qwen3.8-27B shows where many developers think the real adoption wave may come from: a smaller open-weight descendant in the same family, possibly inheriting some of the flagship’s post-training or distilled capabilities @kimmonismus @TheZachMueller

This is the key split in the open-model story: ecosystem influence and benchmark legitimacy come from releasing the 2.4T flagship; practical deployment at scale may come from the 27B release.

Licensing controversy and geographic restrictions

The most concrete skeptical reaction was not about performance, but about the license.

OstrisAI flagged what they read as a license prohibition covering the USA, EU, UK, and Korea, saying the terms appeared to forbid even downloading the model from the US @ostrisai

That concern echoed a broader discussion happening simultaneously around another open-weight release, MiniMax H3, where users argued that geographic restrictions undercut claims of openness @kimmonismus

No clarifying Qwen license tweet appears in this dataset from Alibaba itself, so the restrictive-license reading remained unresolved within these tweets

For engineers, this matters more than the marketing label. “Open weights” can still mean:

no OSI-style open-source rights,

use-case restrictions,

export/jurisdiction limits,

or no legal permission for commercial deployment in key regions.

That licensing ambiguity is one of the main reasons some of the reaction was more cautious than celebratory.

Why the launch matters strategically

This was widely read as a strategic shift by Alibaba, not just a routine product update.

ZhihuFrontier explicitly framed the move as Alibaba choosing ecosystem influence over exclusivity, arguing that earlier Max models stayed closed while the open line had previously topped out around Qwen3-235B @ZhihuFrontier

In that reading, DeepSeek, Kimi, and other Chinese open models weakened the premium of keeping top-tier systems API-only, pushing Alibaba to compete on ecosystem adoption as well as model quality @ZhihuFrontier

Multiple observers connected Qwen3.8-Max to a broader Chinese-model surge:

“Top three spots in front-end design are now shared between two Chinese and one Western model” @kimmonismus

“Remember when China was 2 years behind?” @matvelloso

“The open weights frontier has been consistently dominated by labs from China for the last two years” @_micah_h

Some posters escalated this into a geopolitical concern that US labs cannot rely on closed-model leads forever, especially if Chinese labs keep pushing frontier-ish systems into open-weight channels @kimmonismus

A subtext here is that the moat may be shifting:

not just raw pretraining,

but post-training, agent harnesses, inference infra, distillation pipelines, and developer lock-in.

That is exactly why an open-weight flagship at 2.4T is strategically valuable even if relatively few teams ever self-host it.

Model architecture and sparsity implications

The technical profile suggests Alibaba is leaning harder into sparse MoE than some rivals.

If the 95B active / 2.4T total number quoted by ZhihuFrontier is accurate, Qwen3.8-Max activates only about 4% of total parameters per token @ZhihuFrontier

ZhihuFrontier contrasted this to Qwen3-235B-A22B, which they say activates closer to 10% @ZhihuFrontier

Elie Bakouch’s broader comment—“the two biggest OSS models in the world use linear attention?”—captures another architectural thread in the ecosystem conversation, though it was not directly tied to Qwen3.8-Max with a cited source in-thread @eliebakouch

The wider thread around sparse MoE and Switch Transformers reflects why people care about these parameter numbers: frontier open models can look “huge to store yet still cheap to run” by only activating a narrow expert slice per token @ProfTomYeh

This is likely part of how Alibaba can cut API pricing while scaling total parameter count upward: bigger expert pool, lower active footprint, lower effective inference cost, assuming routing and systems optimizations hold up in production.

Long-horizon agents, cowork, and benchmark fit

Qwen3.8-Max was pitched less as a chatbot and more as a model-harness substrate for long-running work.

Alibaba’s own language emphasized “coding and cowork” rather than generic assistant use @Alibaba_Qwen

The launch claims map unusually well to the current “long-horizon agents” discourse:

10+ day autonomous coding

500+ turns in chip optimization

365-day business strategy @Alibaba_Qwen

ZhihuFrontier’s benchmark picks—PaperBench, CoWorkBench, WideSearch—all emphasize persistent objective maintenance, tool use, and trajectory coherence rather than one-shot Q&A @ZhihuFrontier

Omar Sar0 explicitly linked the release to agent harnesses, saying using Qwen3.8-Max in Hermes Agent makes it hard to deny how much open frontier models have closed the gap with closed frontier systems @omarsar0

Cline’s separate thread about open-weight models is relevant context: they argue many open models are RL-trained to spend more tokens on verification and work best when the harness lets them lean into that behavior, producing ~20% gains from harness changes alone @cline

That fits Qwen3.8-Max’s launch narrative unusually well. The implication is not simply “model is smarter,” but “model may be especially competitive when paired with a harness designed for long-running verification-heavy work.”

Different perspectives in the reaction

Supportive

Strong enthusiasm from open-model developers and infra providers:

“Qwen 3.8 Max and a new local 27B Qwen 3.8 is coming” @Teknium

“Yes, we will have Qwen3.8-Max” @baseten

“Try Qwen3.8-Max on Hermes Agent…” @omarsar0

“Nice! An open source max model” @NerdyRodent

Several commenters treated the release as proof that open models are at or near frontier parity on meaningful workloads @JonathanRoss321 @kimmonismus

Neutral / analytical

Jamin Ball’s thread was the main “yes, but” reaction:

pricing gap may be overstated,

token efficiency matters,

infra burden remains extreme for >2T open models @jaminball

Nrehiew questioned whether performance gains might come disproportionately from post-training rather than novel pretraining, essentially asking how much of the delta is recipe vs scale @nrehiew_

Vals added an important methodological note: Alibaba’s reported Terminal Bench results modify benchmark timeouts, whereas Vals preserved original timeouts @ValsAI

Skeptical / opposing

License concern was the clearest substantive criticism: if usage is restricted in major markets, “open” becomes a narrower claim @ostrisai

Some of the strongest skepticism was indirect: if these giant open-weight models require supernodes and careful harness engineering, then their practical competitive effect may be less dramatic than leaderboard headlines suggest @jaminball

There was also broader ecosystem skepticism that benchmark jumps alone prove full parity with the strongest closed models; e.g. some users argued open source is “very close” but not actually there yet on top-end agentic coding @scaling01

Context: Qwen3.8-Max inside the 2026 open-model cycle

The launch sits in a dense cluster of giant open or quasi-open releases from Chinese labs.

The comparison set repeatedly mentioned in the discussion:

Kimi K3 at 2.8T

GLM-5.2

DeepSeek V4 Flash / Pro

MiniMax H3 on the multimodal/video side @jaminball @kimmonismus

Artificial Analysis commentary cited in-thread said Chinese frontier models have generally trailed top US models by about 3–9 months, while the open-weight frontier itself has been dominated by Chinese labs for roughly two years @_micah_h

This helps explain why the release drew such outsized attention: it is not just another model launch, but part of a visible realignment where:

China is strongest in open-weight frontier scale

US labs still often lead in top closed-model performance

the gap is narrowing on select domains like coding, design, and some multimodal tasks @_micah_h @kimmonismus

Practical implications for engineers

For engineers, the most important questions are less about marketing claims and more about deployment shape.

If you want frontier-ish open-weight quality, Qwen3.8-Max suggests the tradeoff space is now:

very strong eval performance

aggressive token pricing

huge serving footprint

possible license/jurisdiction constraints

The 1M context and 128k output numbers make it viable for repository-scale and workflow-scale tasks where transcript reuse and cache pricing matter @ValsAI @Alibaba_Qwen

The cached-token price of $0.25/M is especially relevant for agents repeatedly replaying codebases, tool traces, and large instruction prefixes @Alibaba_Qwen

The announcement of Qwen3.8-27B may be just as consequential as the flagship, because it is the tier likeliest to become actually usable across broader open-source stacks and local-serving ecosystems @Alibaba_Qwen @kimmonismus

Several developers already framed the release in terms of downstream harnesses and agents, not just chat UX: Hermes Agent, Command Code, Baseten, and likely any provider supporting OpenAI/Anthropic-compatible protocols can slot it into existing workflows quickly @Alibaba_Qwen @Alibaba_Qwen @baseten

One notable interpretation from TeortaxesTex was that Qwen 3.8 Max may be:

exceptionally strong on image recognition/labeling

potentially sample efficient

and distillable/OPD-able into Qwen 3.8 27B for task-specific parity, implying a route from flagship capability to laptop-deployable specializations @teortaxesTex

Other Topics

Agent infrastructure, harnesses, and long-horizon systems

A detailed survey summary argued that long-horizon capability is a model × harness property, not just a model property; it breaks failures into goal drift, context corruption, and sparse-reward/irreversible-action issues, and frames the control plane as shifting from prompt engineering to runtime harnesses @ZhihuFrontier

Cloudflare launched @cloudflare/computer, an agent runtime that dynamically routes between isolates and Linux containers so each agent gets “a computer of its own” @Cloudflare

Cursor reported 20–30% better token efficiency for cloud agents and 80% better efficiency on computer-use runs, plus launched plugins for Google Workspace access across Gmail, Drive, Calendar, Docs, and Sheets @cursor_ai @cursor_ai

LangChain signaled managed Deep Agents moving to public beta, with built-in evals, memory, OAuth tool access, channel integrations, and sandboxing @hwchase17

Several posts emphasized that harness choice materially changes benchmark outcomes and production behavior:

endpoint choice changed Kimi K3 results dramatically on CEO-Bench @tonychenxyz

Cline says open-weight models often benefit when allowed to spend extra tokens on verification, yielding ~20% gains in their runs @cline

a new paper organized 41 agent failure modes by interaction edge rather than single component, with automated labeling reaching κ = 0.76 vs humans @omarsar0

Benchmarks, evals, and automated research/post-training

RSIBench-Data results put Kimi K3 + Kimi Code at 27.317% weighted score across six benchmarks, including 50% SWE-bench Verified and 17% SWE-bench Pro @FanqingMengAI

Intology said its automated AI research system Locus is SOTA on PostTrainBench, and that Locus-post-trained Qwen3 1.7B Base variants surpassed the official human post-trained Qwen3 1.7B release; on live Kaggle comps it reached the 4th highest average rank after 16 days @intology

Epoch updated MirrorCode with Claude Fable 5 at 64% solve rate and GPT-5.6 Sol at 20%, using 15 Medium/Large programs, 2 languages each, and 10B tokens per attempt @EpochAIResearch

Shahules argued benchmarks should release trajectories, not just scores, because task defects and brittle verifiers can dominate failures; they also highlighted ITSMBench as an open benchmark with trajectories @Shahules786

New eval/benchmark artifacts included:

MerchantBench: 365-day e-commerce simulation with 98,843 product records, 26 tools, score on cumulative net assets @dair_ai

One Layer Deeper: adaptive-computation challenge based on repeated modular squaring @SolidlySheafy

Artifacts Hub / Adoption Dashboard tracking 792 open models, downloads, intelligence, and geography @natolambert

Open models, inference, and systems engineering

Multiple posts stressed the open frontier is now dominated by giant MoEs from China, with Kimi K3, Qwen3.8-Max, GLM, and DeepSeek frequently compared on scale/cost/perf @_micah_h

Databricks claimed #1 Kimi K3 inference speed/latency on Artificial Analysis, quoting 239 tok/s in one post and separate single-node numbers from Casper Hansen of 947 tok/s batch-32 decode and 152 tok/s single-user on a single B300 node @Yuchenj_UW @casper_hansen_

Vikhyat announced Photon 2.0, compiling Moondream, Qwen 3.5, and Gemma 4 into megakernels spanning the full forward pass @vikhyatk

A systems paper thread on TokTier argued tokenization can consume up to 64% of TTFT in cached-agent workloads, with stateful tokenization reducing TTFT by 16–34% and incremental repair 437× faster than HF tokenizers in some settings @omarsar0

DSPy 3.3.0 shipped:

dspy.Flex for optimizing code + prompts

ReActV2 with native/parallel tool calling

typed provider-neutral LM interface @isaacbmiller1

Multimodal, video, and vision models

MiniMax H3 dominated discussion outside Qwen:

described as a 33B video model with text/image/video/audio references, up to 15s clips, runnable on one RTX 5090 with ComfyUI stack around 40GB and 5s generations in ~5.5 min in early tests @kimmonismus

later ranked #1 open model in Video Arena, +280 pts over next-best open, and tied near the top overall in image-to-video @arena

There was an active license debate around H3 too: one side said it cannot legally be used in the US/EU/UK/Korea under the public license @kimmonismus, while another clarified formal authorization is available via MiniMax and that “cannot legally be used” is too strong @VictorSuOrtiz

Jina released jina-reranker-v3.5, a 0.6B listwise reranker scoring 63.20 nDCG@10 on BEIR, beating Qwen3-Reranker-4B at roughly 7× fewer parameters @JinaAI_

Qwen3.8-Max also drew attention for vision/object detection use cases, including documents, infrared, satellite, and crowded scenes, with claimed per-image cost around $0.007 @skalskip92

Frontier labs, policy, safety, and competition

A large meta-thread in the timeline concerned US vs China and whether Chinese labs are catching up or already ahead in some

来源说明

当前保存的是 RSS 或来源节选，不代表原文全文。请以原始来源为准。

「要点解读」由 AI Stack 依据上方已保存内容整理，不代表来源的完整表述；标注「推测：」的判断来自编辑，不是来源陈述。