2026-08-04
[AINews] Qwen 3.8 Max(2.4T) and 27B, new open weights models for Coding and Cowork
Qwen is so back!
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.
---
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
Không có nhận xét nào:
Đăng nhận xét