AesCode 32B vs North Mini Code: The Weight File Is Not the Commitment

Both of these models are Apache 2.0, both are downloadable, and both were built for a narrow job — and that is where the resemblance stops being useful. AesCode 32B is Microsoft's research checkpoint for generating editable HTML and CSS: slides, posters, dashboards and reports. Its Hugging Face repository was created on September 29, 2026 and the commit titled "Release AesCode-32B" landed October 7, 2026; Microsoft has published no announcement, and the model card carries no release date at all. North Mini Code is Cohere's open-weight agentic coding model, a 30B mixture-of-experts with 3B active parameters, released in June 2026, aimed at software engineering and terminal work. The obvious way to compare them is on benchmarks, and it does not work, because the two have never been measured on the same task and only one of them has been measured by anybody outside its own lab. The comparison that does work — the one that decides whether you can actually run either — is the weight file.

Two checkpoints, two very different bills

This is the axis most comparisons skip, and for two self-hosted models it is the axis that matters most.

• Parameters — AesCode 32B: 33B, dense. North Mini Code: 30B total, 3B active, mixture-of-experts.

• Memory to run it — AesCode 32B: the card advises planning roughly 65 GB of accelerator memory for the bf16 parameters alone. North Mini Code: quantized, around 20–24 GB is enough, and the reference deployments include a Mac Studio.

• Serving shape — AesCode 32B: a vision-language model, so the serving path carries images as well as text, and its own example uses four-way tensor parallelism. North Mini Code: text in, text out; the benchmarks used a single terminal tool.

• Licence — both Apache 2.0, weights included. This row is a tie and is not the differentiator people assume it is.

• Context — AesCode 32B: 24,576 tokens in the reported configuration, with training-time prompts and responses capped at 8,192. North Mini Code: 256K in, 64K out.

• What it emits — AesCode 32B: a complete HTML document, tables as HTML tables and charts as ECharts specs. North Mini Code: whatever text or tool call the harness asked for.

• Independent evaluation — AesCode 32B: none published; all figures are Microsoft's. North Mini Code: scored on Artificial Analysis, which also flagged its verbosity.

Read the second row twice. A 32B dense vision-language model that wants 65 GB and four-way parallelism is a different procurement decision from a 3B-active MoE that fits, quantized, in a workstation. If your reason for considering open weights was "we want this on our own hardware", that row is most of the answer before you run a single prompt.

The licence is not the reason to pick either one

Nearly every open-weights comparison opens on licensing, and here it is a dead end: both are Apache 2.0, both ship full weights, and neither carries a usage restriction that would stop a commercial deployment. There is no data-sharing clause to weigh against a discount, no contributor tier, no field-of-use carve-out. Whatever you are choosing between these two on, it is not permission.

What is left after you take licensing off the table is what the model was trained to do — and these two trained in opposite directions.

One was trained on a terminal, the other on a canvas

North Mini Code's model card names its target outright: code generation, agentic software engineering and terminal tasks. Cohere trained it across several agent harnesses — SWE-Agent, mini-SWE-Agent, OpenCode, Terminus 2 — specifically so that it works inside whichever one you already run, and evaluated it on SWE-Bench Verified, SWE-Bench Pro, Terminal-Bench v2 and Terminal-Bench Hard. Its published weakness is a real one and was found by an outside evaluator rather than volunteered by the vendor: it emits roughly three times the output tokens of its size-class median. Verbose models write long agent trajectories, and long trajectories are what you pay for once the promotional pricing ends.

AesCode 32B was trained on something else entirely and the card is unusually clear that the target is the artifact. The whole point of the design is that a code model cannot see how layout, hierarchy and colour come together on a canvas, so AesCode uses an image generated from the same prompt as an aesthetic reference while keeping the prompt as the source of truth for content. Supervision is a "design graph" of nodes and edges, seven reward channels — execution, text, boundary, table/chart, layout, whitespace, design — six of them deterministic verifiers and one a vision-language judge working from a per-sample rubric. Candidates are rendered in a sandboxed Playwright browser with external requests blocked and scored from the DOM, computed styles, bounding boxes and a screenshot.

That machinery exists to make one thing true: the output is a rendered page that a human will open and edit, not a patch applied to a repository. AesCode 32B has never been trained to use a shell, and nothing about the reward design would reward it for trying.

The evidence asymmetry, stated honestly

North Mini Code has been run by a third party. Artificial Analysis scores it 33.4 on its Coding Index and 27.6 on its Intelligence Index, and publishes the verbosity finding. Cohere's own headline figure — 61.0% pass@1 on SWE-Bench Verified — is vendor-reported but checkable against a model anybody can download, and the card is candid about its harness choices, going as far as noting that for Terminal-Bench Hard it followed Artificial Analysis's own methodology to keep the comparison fair.

AesCode 32B has no equivalent. Its 300-sample infographic evaluation, three generations per prompt with no selection, produces 85.89 Overall against 81.28 for GPT-5.5 and 80.39 for Claude Opus 4.8 — all measured inside Microsoft's harness, on Microsoft's samples, with the scoring channels Microsoft trained the model against. There is a real chance the numbers are fair; the design is more transparent than most vendor tables, the verifier stack and rubric builder are open-sourced, and the card publishes the fact that no model in the comparison clears 60 on its style dimension. But no outside lab has reproduced any of it, and there is no arena listing, no leaderboard placement and no independent throughput test. If a vendor benchmark that happens to favour the vendor's own benchmark — an infographic task — is doing a lot of work in your decision, you are being asked to take the only reading available.

Note what this means for the matchup specifically: there is no shared number. None. One model is scored on whether software engineering tasks pass, the other on whether text is legible and layout does not overflow the canvas. Any page that puts 33.4 next to 85.89 is comparing a Coding Index to an infographic Overall score and neither model knows what the other was tested on.

Neither one is a general-purpose model, and that is the real cost

Here is the trap in both halves of this comparison. Adopting either model means maintaining a self-hosted serving path for a specialist: 65 GB of VRAM and a multimodal stack for the one that renders pages; a slot for a 3B-active MoE for the one that handles terminal tasks. Both are good at their job and neither can do the other's, which means a team that wants both capabilities is now running two inference deployments — and for everything in between, the model that is neither a page renderer nor a terminal agent, they are running a third.

That is the case where a routing layer earns its place rather than being bolted on. One key over 200-plus models at provider list price, no markup, means the specialist work stays on your own hardware where the economics and the data handling make sense, and the routine work goes to whatever hosted model is best for it that week. The Qwen3-VL backbone family is a good illustration of how thin the line gets: Microsoft built AesCode 32B on Qwen3-VL-32B-Instruct, and vision-language members of that family are callable on the OrcaRouter catalogue — Qwen3-VL 235B A22B Instruct at $0.40 per million input tokens and $1.60 output, or the smaller Qwen3-VL 8B Instruct at $0.18 and $0.70. Neither of those is AesCode 32B: the fine-tune is Microsoft's and no provider serves it. But if what you wanted from AesCode 32B was a model that reads a screenshot and writes code, and you did not specifically need the aesthetic-design fine-tune, the callable option exists and costs a fraction of the GPU. Automatic failover underneath the same endpoint is the difference between trying that and betting a production path on it.

How to actually decide

Take the licence off the table first, because it is a tie and it is not going to decide anything.

Then ask what you want the output to be. If the answer is a rendered, editable page — a deck, a report, a dashboard — North Mini Code cannot help you and AesCode 32B is the only one of the two pointed at the problem. Go in with clear eyes about the 65 GB, the 24K context, the absent independent evaluation and the fact that this is an unannounced model with a double-digit download count. It is a research release, and research releases reward teams who can afford to be curious.

If the answer is a change to a repository — a migration, a fix, a multi-step terminal job — North Mini Code is the one with the harness training, the 256K window, the 3B-active deployment profile and an outside measurement of both its quality and its verbosity. Budget for the token count, not the headline rate.

And if your list of things to do this quarter contains both a design artifact and a repo migration, do not buy this as a two-model decision when it is really a six-model decision. Run the specialist where it pays, route the rest.

The unannounced-model problem

One final thing separates these two, and it is not technical. North Mini Code has a vendor behind it that published a model card, a blog post, a Hugging Face Space to try it in, and a paper trail of benchmark methodology you can argue with. AesCode 32B has a repository, a README and a citation line that reads "Under review, 2027". Microsoft has not said it exists.

That changes when you can adopt it. With North Mini Code, the open question is whether its verbosity makes it uneconomic at scale — a question you can answer by running it. With AesCode 32B, the open question is what Microsoft intends to do with it: whether this is a research artifact, the first release of a product line, or something that will be quietly replaced by a 70B version in six weeks. Nothing in the repository answers that, and no amount of reading the model card will. Watch for three signals — an official announcement, an independent reproduction of the 85.89, or an inference provider picking the checkpoint up — and treat the absence of all three as the current state of the evidence rather than as a reason the model is uninteresting.