AI Dev Weekly is a Thursday series covering the AI developer news that changes what you can build, what it costs, and how you should operate it.
Last week, Kimi K3.1 Pro was a hosted model with a promised weights release. This week, there's a checkpoint you can inspect. Tencent Hy4 took a different route into developers' hands through a gateway, while OpenAI added new image outputs and Sarvam released a model that scores choices instead of writing answers.
The useful question this week is what you actually get back: a checkpoint, a hosted response, an editable asset, or a probability your application can act on. Those are four different integration jobs.
1. Kimi K3.1 Pro weights are downloadable
Moonshot's Kimi K3.1 Pro repository now hosts the checkpoint promised in last week's edition. The model card specifies a mixture of experts with 2.8 trillion total parameters, 104 billion active parameters, and a 1,048,576-token context window. The published model uses MXFP4 weights and MXFP8 activations and accepts text and images. Moonshot publishes a model-specific license, which is the document to check before redistributing it or offering it to customers.
Active parameters describe which portion of the model participates in processing a token. They are not a download size or a serving memory budget. Even a back-of-the-envelope calculation at four bits per parameter comes to 1.4 TB for 2.8 trillion parameters before accounting for the actual weight layout, runtime buffers, or caches. That's an illustration of scale, not a measured hardware requirement. The advertised context window also doesn't tell you how much context you can afford at your target concurrency.
My take: Try the hosted route first if you need a quality baseline, then price self-hosting against that baseline. Use the same repository tasks and measure successful patches, review effort, and wall-clock time. If you do deploy the weights, record the checkpoint revision and runtime alongside your results. Our K3.1 Pro API guide covers the hosted side, and the GPU memory planning guide explains the distinction between weight memory and total serving cost.
2. Tencent Hy4 arrives on Vercel AI Gateway
Vercel added Tencent Hy4 on November 3, under the model ID tencent/hy4. Its listing describes a 1M-token context window and support for image input, reasoning, tool calling, and structured outputs. Vercel says the gateway route is 50% off through December 31, 2026. That's a promotion on the listed route, not a permanent inference rate.
The listing is for Hy4, not the earlier Hy4 Preview checkpoint. Seeing the newer model in a gateway picker confirms that you can call it there; it doesn't establish that the same checkpoint is downloadable. Likewise, support for a tool-call interface doesn't mean the model arrives with access to your browser, shell, or repository. Your application or coding client still supplies those capabilities and sets their permissions.
My take: Test Hy4 on something that needs both visual input and a constrained action, such as inspecting a screenshot and proposing a change against a known codebase. Check what happens if the image is ambiguous or the requested tool is unavailable. Separate the trial price from the budget you would approve after the promotion ends. The API pricing comparison and model fallback patterns are useful when adding it as another route rather than replacing your default.
3. GPT Image 2.6 adds SVG, GIF, and more explicit edit controls
OpenAI's November 4 API changelog introduced GPT Image 2.6 Sunburst and Flare. The image generation guide positions Sunburst for detailed visual work and Flare for faster iteration. Both support SVG and animated GIF output through the Image API. The new complexity setting controls planning effort separately from rendering quality; it accepts low, medium, or high and defaults to high.
Markup-guided editing is also available on supported 2.5 and 2.6 models through the Image API. You can send an annotation layer to show the region you mean and explain the desired change in the prompt. That's useful for a reviewer circling a product detail or a sign, but don't assume unmarked pixels will be preserved exactly.
My take: Test a pipeline, not a pretty sample. For an SVG, verify it renders correctly in the target browser and sanitize it before serving it as active content. For a GIF, check size, legibility, and playback in the places customers will see it. For an edit, compare the rest of the image with the original as well as reviewing the requested change. If a fast iteration requires repeated retries, compare cost per accepted asset rather than assuming the model label settles the economics. See our image API pricing guide for the broader cost framework.
4. Sarvam Prism makes the output a decision, not a paragraph
Sarvam released Prism on November 3, a decision model built on Sarvam 105B, with Apache 2.0 weights. You submit the text or JSON state and typed questions; the response gives a distribution over the options rather than generating a textual answer. The API reference covers yes/no, fixed-choice, and ordered-scale questions. It does not currently accept images.
This is a concrete alternative to prompting a chat model to return a label and then parsing the reply. Say your support team has four queues. Give the model the ticket and those four choices, then make your application decide whether the result is strong enough to route automatically. A probability is evidence, not authorization. If a false positive is costly, use a higher threshold or send that case to a reviewer. Check those thresholds on your own labels and languages rather than borrowing a number from a vendor evaluation.
My take: Start with shadow routing. Record the predicted queue alongside the human assignment without changing who receives a ticket. Review the errors by queue, not just the overall accuracy. A model can look good on the common cases and still fail on the rare escalation that matters most. Our regression testing guide covers how to preserve a useful test set as prompts and models change.
Quick hit: GPT-6 Rosalind is a restricted research release
OpenAI's October 30 changelog entry lists gpt-6-rosalind for specialized life-sciences reasoning and scientific tool use. Access is limited to approved organizations and authorized internal research users. Treat that as a separate access program, not a general-purpose GPT-6 upgrade for ordinary API projects.
What I would test this week
Choose the release that changes a job you actually have. Compare Kimi hosted and self-hosted only after defining the throughput and data requirements. Run an image edit and inspect what changed outside the marked region. Or take a real queue, collect ground truth, and see whether a decision endpoint can reduce manual routing without missing escalations. Each test should tell you what to do next even if the new model loses.
FAQ
Can I download Kimi K3.1 Pro now?
Yes, Moonshot hosts the checkpoint on Hugging Face. That does not make it a single-workstation model; evaluate the complete serving footprint before provisioning hardware.
Is the Hy4 gateway discount permanent?
No. Vercel's announcement specifies a 50% discount through December 31, 2026.
Are complexity and quality the same setting in GPT Image 2.6?
No. Complexity controls planning effort. Quality controls rendering. Treat them as separate variables when comparing turnaround and output.
Does Sarvam Prism write explanations for its choices?
Its decision endpoint returns typed choices and probabilities, not generated explanations. If you need prose for reviewers, keep it separate from the decision record.
Originally published at https://www.aimadetools.com
Top comments (0)