AI & Tech Daily
Qwen4-Max and the Cost of Getting AI Work Done
Qwen4-Max brings vision, a million-token context and published API pricing. Jesse examines the cost of a finished task, Tencent's smaller active-compute model, NVIDIA's science-agent package and Oak Ridge's operational Lux supercomputer. Also: US controls on remote computing access, an exploited Check Point gateway chain, Microsoft's next local Windows model, and the access and execution defaults that developers can act on now.
Full transcript
Read the episode.
I'm Jesse Owen. This is AI and Tech Daily.
Qwen4-Max and the Finished Task
A million-token context window can hold a lot of work. With Alibaba's new Qwen4-Max, the more interesting number is what it costs to get that work finished.
Our main story today is Qwen4-Max, released on 16 November, and how its combination of vision, tools and pricing changes the options for developers building document-heavy AI applications.
Alibaba specifies native vision and a context window of one million tokens. Tokens are the chunks of text a model handles; context is the material it can work with during a request. The model accepts text, images and video, and produces text. It also supports tool calling and structured outputs.
Those features fit together in a fairly practical way. Visual input gives an application a route to work with material whose meaning depends on what appears on the page or in a frame. A large context window provides room for supporting material. Tool calling lets the model request actions from software around it, while structured output gives that software an agreed shape for the response.
Consider a developer building a visual-document workflow. Reading the document is one part of the job. The application also needs to extract useful information, pass it into another system and handle the result. Qwen4-Max brings several of those building blocks into one hosted model. How well they work on a particular collection of documents remains a question for that application's evaluation.
The published prices are one US dollar fifty per million input tokens and four dollars fifty per million output tokens. Cached input is seven and a half US cents per million tokens. That much cheaper category applies to input served from cache. Ordinary input has its own, higher rate.
That distinction affects the economics of repeated work. An application that can reuse cached material has a different cost profile from one constantly supplying fresh input. The amount of text generated also matters, because output is priced separately. So a long context window gives you capacity, while the actual amount you send and receive contributes to the bill.
Independent evaluator Artificial Analysis gives Qwen4-Max an Intelligence Index score of 43 and measures a cost of three US dollars twenty-seven per benchmark task. The task figure is useful because it puts a workload around the token prices. It describes the evaluator's workload, so I'd use it as a comparison point alongside results from the work an organisation actually wants done.
For me, the useful buying unit is an acceptable completed task. A low token rate can be attractive, but a workflow that needs several attempts consumes more input, more output and more time. Extra reasoning can also change the amount of work a model does before it returns a usable result.
That gives developers a concrete comparison to make: how much does it cost to get an answer that the application can use, including the retries? A document extraction task and a coding task can put very different demands on the same model. Keeping those workloads separate makes the price comparison more informative.
There's also a deployment choice here. Qwen4-Max's weights aren't publicly available, so this release gives developers a hosted model rather than something they can download and run on their own hardware. For organisations comfortable with that arrangement, the appeal is a broad set of capabilities without taking on model hosting. The strongest case for adopting it will come from a useful workflow completed at a sensible cost.
Tencent Sizes AI for Routine Work
With that bigger model picture in mind, Tencent's new release takes a more focused approach to everyday work.
Tencent launched Hy4 flash on 16 November through TokenHub and applications including WorkBuddy. It's aimed at document work, presentations and other everyday agent tasks.
Tencent specifies multimodal input and up to a million tokens of context. The model has 224 billion parameters in total, with 12 billion active. Parameters are the learned values inside a model. The active figure describes the portion brought into use during computation, which helps explain why a model with a large total size can still target lighter workloads.
The advertised prices are approximately seven US cents per million input tokens and twenty-eight point one US cents per million output tokens. Tencent's speed and task-cost comparisons come from its own testing. The immediate offer is a low-priced option within the platforms that currently support it.
I find the division of labour more interesting than the price alone. Tencent directs complex, long-running software engineering towards the larger Hy4. Flash has a more everyday brief. That gives a developer an explicit reason to distinguish a document task from a difficult engineering job when choosing where to send the work.
For example, preparing material for a presentation and carrying a complicated software change through several stages place different demands on an agent. Treating both as the same workload can make it harder to see where the money goes. Separating them creates room to use a smaller active-compute option where it does enough, while reserving the larger model for the more demanding work.
My reading is that this could make routine automation cheaper to operate for teams already using Tencent's supported platforms. The useful question is whether flash completes those routine jobs well enough to keep them off the larger model.
Access is still platform-specific. OpenRouter integration and an open-source release are planned steps. There's no verified downloadable release in this announcement, and Australian account eligibility hasn't been established. That leaves current users with a hosted option to assess, while developers who need to run the model themselves are waiting for a separate release.
A Chemistry Toolkit for Science Agents
Let's leave office work for a moment and look at what a more specialised AI package gives researchers.
NVIDIA introduced Nemotron for Science on 16 November. It combines the Nemotron 3.5 Super VL model with chemistry data and post-training recipes for domain-specific research agents. Post-training is the additional adaptation used to shape a model for particular tasks after its initial training.
The chemistry dataset, LitmusBench version zero point two, contains 47,118 training questions and 2,056 test questions about chemical properties and experimental data. Those separate sets give researchers material for adaptation and a set of questions against which to assess the result.
That separation is useful. If you're changing a model to work in a specialist field, you need a way to see whether the changes help on questions beyond the ones used to teach it. A published training and test split gives different research groups a shared starting point for that work.
The package also brings the model, data and adaptation recipes together. For a research organisation, that can reduce the amount of groundwork required before it can investigate a particular scientific use. A researcher can examine the questions, adapt the model and evaluate its responses within a defined subject area.
NVIDIA says the model and dataset are available on Hugging Face. It describes an optimised local deployment on a DGX Spark with 128 gigabytes of memory. That's a substantial hardware requirement, and it puts a concrete shape around the local option: a research organisation can bring the work onto its own machine if it has the appropriate system.
The data permits commercial and non-commercial use, with different licences applying to different source material. Questions derived from ChEMBL use the Creative Commons Attribution ShareAlike four point zero licence; the other questions use OpenMDW one point one. For anyone building on the package, that split is part of the reuse terms.
I'd put the value in the reusable research process this enables. An organisation can adapt a model around chemistry questions and measure how it behaves, with the ingredients available for inspection. That makes specialist development more concrete than choosing a general assistant and hoping it understands the field. The published questions provide a place to begin testing which scientific tasks deserve a more specialised agent.
Lux Opens for Scientific Work
A model package is one ingredient; researchers also need somewhere to run the models and the simulations around them.
Oak Ridge National Laboratory announced on 16 November that its Lux supercomputer has opened to scientific users. Lux adds operational AI and simulation capacity for the US Department of Energy's Genesis Mission.
The machine combines AMD Instinct MI355X GPUs, EPYC processors and ROCm software in an HPE system. More useful than the component list for many researchers is the mix of work it can accommodate: model training, inference and agent workflows alongside traditional simulation.
Simulation calculates how a system behaves using a computational model. AI training and inference involve different kinds of work, and a research programme may need both. Putting those capabilities into an operational research system gives selected projects a place to develop workflows that use them together.
Lux supports both Slurm and Kubernetes for scheduling. These manage how work gets assigned to computing resources. Supporting both gives research groups options for bringing different styles of workload onto the machine, instead of making every job fit a single scheduling approach.
The allocation is also significant. Lux has approximately three and a half million node-hours annually, with roughly half supporting the Genesis Mission and half available to AMD. A node-hour measures one computing node used for one hour. It's a way to describe the amount of machine time available to divide among projects.
This is allocated institutional capacity. Access runs through those arrangements, and the announcement doesn't establish general availability for Australian researchers.
Named programmes include fusion-plant design, materials development and longer-range water prediction. These are uses for the newly available capacity, with the research results still ahead.
For research leaders, I'd pay close attention to allocation and workflow support. A project needs enough time on the machine, a way to schedule its jobs and support for the software it uses. Lux becoming operational turns those questions into practical planning decisions for selected users. The opportunity is to put AI and simulation to work within the same research programme, with actual computing time behind it.
US Controls Reach Remote Computing
There's a separate development for organisations renting advanced computing across borders, and it concerns who gets access.
On 16 November, the US Bureau of Industry and Security announced an interim final rule addressing remote access to advanced chips and servers. The scope includes computing supplied through data centres in third countries.
BIS describes the measure as aligning controls on remote computing access with restrictions on physical shipments of advanced technology. The consequence for procurement is that rented computing capacity comes into the picture alongside the sale and movement of hardware.
A server can remain in the same building while people elsewhere use its capacity. That's the arrangement the announcement brings into focus. For an affected cloud provider, the location of the equipment is one part of the question; the people and organisations receiving remote access are another.
My reading is that international compute procurement becomes more dependent on understanding the access arrangement. A customer buying capacity through a provider needs to consider how that arrangement fits the controls, even when the customer never buys a chip or takes delivery of a server.
The official announcement doesn't give an effective date or a complete implementation timetable. Commencement and transition arrangements therefore remain open questions for affected contracts. That timing is the specific detail providers and customers still need before they can work out when their arrangements may have to change.
Check Point Gateway Intrusions
Back inside the network, a security gateway investigation shows how an intrusion can spread beyond the appliance itself.
Check Point disclosed on 12 November that attackers had chained PHP2Shell vulnerabilities with a local privilege-escalation flaw in Gaia to obtain root access on unpatched gateways. Root access gives an attacker the highest level of control over the affected system.
The company dates the observed activity from 30 October to 4 November and reports a couple of dozen confirmed compromised systems in one region. Its investigation documented a pivot from a gateway to an ESXi host and the use of destructive malware.
The chain helps explain the response. The Gaia flaw requires an existing local foothold. Once an attacker has that foothold, privilege escalation can increase the control available to them. The documented movement to an ESXi host then shows why an investigation may need to follow activity beyond the original gateway.
For an organisation operating one of these systems, the gateway is part of the route into the wider environment. My takeaway is to treat a suspected compromise as an intrusion investigation with a network dimension. The affected appliance provides the starting point for understanding where the attacker went.
The update path is release-specific. CERT Santé lists fixes for the supported Gaia release families R82.20, R82.10, R82 and R81.20 at particular Jumbo Hotfix levels. Operators need the fixed level for the branch they're running. The older R81 and R81.10 branches are unsupported, and the advisory lists no fix for them.
Product boundaries also matter here. Spark and Smart-1 Cloud are outside the listed scope of the Gaia privilege-escalation flaw. Spark still needs separate consideration under the PHP2Shell advisory. A product's exclusion from one part of this chain doesn't settle its exposure to the other part.
Check Point's guidance combines the appropriate updates with a compromise investigation. Updating closes the relevant vulnerabilities; investigation establishes what happened during any earlier access. That distinction is especially useful with a gateway, because the documented attack continued into another system. For affected operators, the work is to apply the release-specific fix and follow the evidence of access through the environment.
Windows Prepares a Larger Local Context
On the desktop, Microsoft is preparing a model change that gives local AI applications much more room for input.
Microsoft announced Aion Instruct on 13 November as the planned replacement for Phi Silica, alongside support for additional GPUs. Aion has 2.3 billion parameters and a 32,000-token input context. Phi Silica has 3.8 billion parameters and a 3,500-token input context.
So the planned model is smaller by parameter count while accepting a substantially larger input. For a developer working with longer material locally, that creates room to evaluate tasks that need more text in the same request.
A standalone evaluation package is available now. Microsoft's current targets are a Windows Insider rollout on 4 December, followed by retail rollout in January 2027.
The migration also needs attention to model-dependent behaviour. Custom LoRAs, which are lightweight model adaptations, specialised prompts and output parsers can be affected. A parser is the code that turns a model's response into something the application can use. Microsoft is also introducing invisible text watermarking.
I'd treat the evaluation package as a chance to check the application's behaviour before the retail transition. An API can stay familiar while the model behind it responds differently. The larger context is an opportunity to handle more material on the device; the practical development work is making sure existing adaptations and response handling still fit.
What Changes for You
If you're deciding what you can use now, there are two access details worth keeping close to hand.
For Qwen4-Max, developers can use the paid API through QwenCloud. Application users can access it through QwenStudio or QwenWork. That gives builders and people using those applications a route to the new model without setting up their own model-serving infrastructure. It's a hosted choice, with usage charges for the API.
For existing browser-agent builders, Browser Use 1.0 brings a more immediate migration decision. Released on 10 November, it replaces the earlier agent architecture with persistent Python execution directly over Chrome DevTools Protocol, the interface software uses to control and inspect Chrome. Persistent execution lets the agent work in an ongoing Python environment as it interacts with the browser.
The default browser has moved from local Chromium to Browser Use Cloud. Local execution is still selectable, and the default model uses Browser Use's gateway. If your existing application relied on the browser running locally, that location now needs to be an explicit choice when adopting the new release.
It's a breaking migration from the zero-series releases, so existing configuration needs attention. There's also a billing detail that's easy to miss: the reported cost field excludes cloud-browser and proxy charges. Those costs sit outside the amount displayed there.
I'd make the execution location and full bill the first migration decisions. They determine where the browser work happens and what operating it costs. A builder who wants local execution can still choose it; someone adopting the cloud default needs to include the browser and proxy charges in the application's budget.
You'll find the sources and full transcript at owenonthenet.com. Thanks for listening.
Sources
Reporting behind this episode.
- alibabacloud.com/en/press-room/alibaba-unveils-qwen4-max
- qwencloud.com/models/qwen4-max
- artificialanalysis.ai/models/qwen4-max
- tencent.com/tencent-launches-hy4-flash-to-power-high-speed-cost-efficient-ai-deployment-for-everyday-workflows
- nvidianews.nvidia.com/news/nvidia-introduces-nemotron-for-science-to-accelerate-discoveries-with-agents
- huggingface.co/datasets/nvidia/Nemotron-RL-litmus-bench-v0.2/raw/main/README.md
- olcf.ornl.gov/2026/11/16/lux-genesis-mission
- bis.gov/press-release/bis-issues-rule-addressing-remote-access-american-super-intelligence-technology
- blog.checkpoint.com/security/security-advisory-active-exploitation-of-php2shell-and-local-privilege-escalation
- research.checkpoint.com/2026/php2shell-exploitation-on-check-point-gateway
- cyberveille.esante.gouv.fr/alertes/check-point-cve-2026-123039-2026-11-13
- github.com/browser-use/browser-use/releases/tag/1.0.0
- devblogs.microsoft.com/foundry-on-windows/aion-instruct-local-ai-windows-gpu