Perceptron Mk1.5 vs Qwen 3.8: The Robot Needs Both of Them and Neither One Is a Substitute

A robot that has to pick something up needs two things from a model, and no single model in this pairing gives you both. Perceptron Mk1.5 returns geometry: given video frames, it can hand back a point, a box, a polygon or a timestamped <track> element, and its context window is 36,864 tokens. Qwe​n 3.8 — the name that resolves to the vendor's open-weights Qwen3.8-2.4T-A95B core — is a 2.4-trillion-parameter mixture-of-experts reasoner with a native 262K window that extends toward a million, and it is text-only, so it cannot look at the frames at all. One is a perception layer that costs $0.15 and $1.50 per million tokens; the other is a reasoning core that costs roughly thirteen times as much going in and four times as much coming out, and has an independent score of 40 on the Artificial Analysis Intelligence Index while the perception model has no independent score anywhere.

Set side by side as competing products, they lose to each other in opposite directions and the comparison produces nothing usable. Set side by side as the two halves of one pipeline, they explain almost every cost and reliability decision in an embodied stack: where the frames get interpreted, where the plan gets made, and what happens to your bill when the reasoning half writes more tokens than the task needs.

The split that makes this pairing make sense

Perceptron Mk1.5 shipped on September 25, 2026 as the successor to Perceptron Mk1, and its job is narrowly defined. It accepts text, images, video and audio, and it returns text plus optional structured annotation — points, bounding boxes, polygons, clips and tracks. The <track> output carries both a spatial observation and the timestamp it belongs to, so a 60-second clip comes back as positions over time rather than one averaged guess. An asset_idx field lets a single request address several images or videos separately, which is what a reference-frame-versus-live-frame comparison needs. Constrained responses come in two flavours, JSON Schema and regex, and the whole point of both is that the output is typed.

Qwen 3.8 is the opposite kind of instrument. The 2.4T core is a fine-grained MoE — 92 layers, 512 experts per layer with 10 routed plus one shared, and 69 of those 92 layers on linear attention — which is how an architecture that large reaches its long context. Where it is strong is reasoning across a lot of text: complex multi-step analysis, long derivations, agentic planning. Where it is incapable is the input side. The open core is text-only, and its reasoning cannot be switched off: every response opens with a thinking block whether you wanted one or not.

That is not a deficiency in a reasoning model; it is a deficiency in a perception model, and Qwen 3.8 is not one. Put the two in sequence and the shape is obvious — Mk1.5 answers "where is the forklift and when did it move", Qwen 3.8 answers "given that, what should the arm do". Neither question is answerable by the other model.

What each half costs, at the rates that actually bill

• Input — Perceptron Mk1.5 $0.15 per million tokens. Qwen 3.8 $2.00 per million on the hosted build that the independent harness measures, an 88% cache discount applied, which brings a cache hit to roughly $0.24.

• Output — Mk1.5 $1.50 per million. Qwen 3.8 $6.00 per million.

• Cache — Mk1.5's cached input is $0.0375 per million, a flat quarter of its standard input rate. Qwen 3.8's discount is percentage-based but deeper, at 88%, so its cached rate is the cheaper of the two in absolute terms and its uncached rate is more than ten times Mk1.5's.

• Cost per finished task — this is where the ranking inverts. Artificial Analysis puts Qwen 3.8's cost per Intelligence Index task at $2.16, ranked 32nd of 115 in its class and rated four out of four units for cost — cheap per completed task despite expensive tokens, because the model generated 170M output tokens across the index runs against a field that rambles further. Mk1.5 has no equivalent figure published by anyone.

• Context — 36,864 tokens for Mk1.5 against a native 262K for the Qwen core, extensible toward a million with the right serving configuration.

• Reasoning control — Mk1.5 exposes reasoning_effort at high, medium, low, minimal or none and defaults to high. Qwen 3.8 locks thinking on, so you pay for the thinking tokens on every call.

Read that list twice, because the two models invert on cost. Per-token, Mk1.5 is dramatically cheaper. Per finished task on an independent benchmark, Qwen 3.8 is measured as cheap and Mk1.5 is unmeasured. Those two statements are only contradictory if you assume they are running the same job, and they are not.

The evidence runs the other way here

In most of this batch's comparisons the open-weights side is the one with a stack of vendor-reported numbers and the closed API is the one with a third-party page. Here it is reversed, and the reversal is worth stating precisely because it changes what you can and cannot verify.

Qwen 3.8 has independent measurement. The Artificial Analysis Intelligence Index puts the 2.4T core at 40, ranked 5th of 115 models in its class at the time of writing, with an output speed of 39.6 tokens per second — 56th of 115, which is mid-pack and worth knowing before anyone plans an interactive loop around it. Index revisions move between snapshots and the honest way to read any of these figures is as directional, but the point stands: somebody outside Alibaba has run this model and published a number.

Perceptron Mk1.5 has no such thing. There is no Artificial Analysis entry and no arena score for Mk1.5 or its predecessor, so every capability figure in existence was produced by Perceptron about Perceptron's own model. That set is unusually specific, which is a point in its favour — 0.9433 on egocentric hand_box against 0.4467 for Gemini 3.1 Pro in the vendor's own run; EgoSchema 80.40 against 81.20 for the same competitor, a narrow loss on the broad comprehension benchmark alongside a claimed 12% edge on the harder EgoSchema subset at 63.75; 0.647 centre-F1 and 0.628 point-HOTA on Molmo2-Track — but specific and independently verified are different properties, and only the second one tells you what will happen on your footage.

Two cautions on reading Perceptron's table, both of which apply to any quotation of it. The LiveVQA-W figure of 56.0 with tools enabled comes from a four-sample majority vote, while the 53.2 it is set against for Gemini 3.8 Flash with Google Search grounding is not a majority vote; the technique is legitimate and it flatters a score relative to a single greedy pass. And the tracking row lists Qwen3-VL-8B at 0.180 centre-F1 taken from that model's paper, sitting in the same column as measured numbers for a different checkpoint family. A paper figure in a measured column is not a like-for-like row, and it is exactly the kind of line that gets quoted without its provenance.

The 2.4T core is not something you can call from us — but its architecture is

This is the part of the comparison where the honest answer differs from what the model's fame suggests. Qwen 3.8 as a name is widely used to mean the open core, and the open core is a download, not an endpoint: 2.4TB of BF16 weights under Alibaba's custom Qwen3.8-Max licence, published in August 2026. We do not serve it, and there is no provider serving it on our side of the fence today — the catalogue entry exists as an announced, not-yet-available tracking page rather than a live route.

What we do serve is the flagship API built on the same 2.4T architecture. It is live as qwen/qwen3.8-max at $2.00 and $6.00 per million tokens, Alibaba's own rate passed straight through with nothing added per token, carrying the 1M-token multimodal window — text, image and video input — and the same hybrid attention design as the open core. If your reasoning half needs to see anything, that is the route, and it is also the route where a vendor rate change lands on our side the same day rather than at your next billing cycle. Alongside it we serve Alibaba's dense sibling qwen/qwen3.8-27b on our own infrastructure at $0.33 and $2.40 per million, with a 262,144-token window and text, image and video input, for the cases where a 27B reasoner is enough and the 2.4T's index of 40 is more than the task needs.

Wiring the two halves without a second contract

The practical reason this pairing matters more than the benchmark argument is that an embodied stack is already two models, and the integration cost of a third is the thing that quietly kills the design. Mk1.5 is not on our catalogue, so reaching it means the vendor's own API — pip install "perceptron>=0.4.0", key in PERCEPTRON_API_KEY, endpoint api.perceptron.inc. The Qwen half is one key away.

Where a gateway earns its place is at the seam. A routing rule that sends every frame-with-a-question to the perception model and every text-only plan to the reasoning model is a configuration line when both sit behind one OpenAI-compatible endpoint, and automatic failover means a provider incident on either side degrades into a fallback rather than a stalled robot. One API covering 200-plus models with list prices passed through at 0% markup is also the cheapest way to answer the question this article cannot: whether your object-tracking task needs Mk1.5's coordinates at all, or whether a general vision model with a much larger window and an independent score would have been enough. Routing a share of real traffic between candidates and measuring on your own frames costs less than any benchmark study you could commission.

What to do with this, concretely

If you are building perception into a physical system, Perceptron Mk1.5 is the only model in this pairing that answers in coordinates, and that is a capability rather than a score — test the hand-box claim on your own video, set reasoning_effort deliberately instead of accepting the high default, and budget for the 36,864-token window as a hard constraint rather than a soft one. If you are building the reasoning layer above it, Qwen 3.8 is measured where Mk1.5 is not, and its cost per completed task is the figure to plan against rather than the $2.00 headline — with the caveat that its thinking cannot be switched off, so a task that does not need a chain of thought is paying for one anyway.

The teams that get this wrong are the ones that try to make one model do both. A 36,864-token perception specialist will not hold the operational history, and a text-only 2.4T reasoner will never see the frame. The pairing is not a compromise between two products. It is the actual division of labour, and the useful question is only which side of it each of your calls belongs on.