Intern-Decision-0.8B vs Qwen 3.8: 853 Million Parameters and a Closed Set of Answers

Intern-Decision-0.8B is the smallest of three decision models InternLM pushed to Hugging Face on the morning of September 26, 2026, and it is not the fastest of them. That is the first thing worth knowing. On the lab's own measurements the 2B sibling runs marginally quicker — 33.28 ms against 33.98 ms — and scores six points higher on the same seven-suite average, so the usual reason to reach for a sub-billion checkpoint, raw speed, does not apply here. What applies is size: 852,985,920 parameters, 1.71 GB of bf16 weights, small enough to sit permanently in the margin of a machine that has something else to do. Set that against Qwen3.8-Max — the hosted delivery of the 2.4-trillion-parameter sparse mixture-of-experts core the vendor published in August, priced at $2.00 and $6.00 per million tokens — and the two are not competing for the same job. One returns a probability distribution over options you supplied in advance. The other writes.

The short answer: Intern-Decision-0.8B is worth having if you need a closed-set decision scored on hardware you already own, and if 1.71 GB instead of 4.43 GB is the difference between shipping and not shipping. It is not worth having if you want the most accurate small scorer InternLM shipped this week — that is the 4B — or if your answer is not already enumerated in your prompt, in which case neither of these two is the right tool and Qwen 3.8 is the only one of the pair that can do the job at all.

What the repository says, and what nothing else does

Start with the evidence, because there is very little of it. InternLM has made no announcement for this model. No launch post, no paper, no repository description. The Hugging Face card links a demo Space and a GitHub repository, and as of this writing the Space returns 401 and the GitHub link returns 404 — the two places the card points you for more information are closed. Three checkpoints appeared within forty seconds of one another at roughly 05:36 UTC on September 26: Intern-Decision-0.8B, Intern-Decision-2B and Intern-Decision-4B, all carrying the same tags — decision-making, multimodal, structured-prediction — and all fine-tunes of Qwen checkpoints, in this case Qwen3.5-0.8B.

What is knowable from the repository is unusually precise for an unannounced release, because the lab shipped its own inference module alongside the weights. The 0.8B loads through a 16 KB inference.py in the model directory, requires Python 3.12 or newer and torch 2.9.1 with transformers 5.14.1, and exposes a DecisionEngine class you instantiate once and reuse. One Python dict in, one response dict out. What is not knowable: whether this is a finished release or an early push, whether a hosted endpoint is coming, and whether the two checkpoints it sits between are meant to be the same product at different sizes or three different products that happen to share a name. Nobody outside InternLM has run any of them.

The sibling problem: the 2B is faster and better

Here is the part of InternLM's own published table that makes the 0.8B hard to place. Reading the three sizes down the same seven suites:

• Speed — 0.8B: 33.98 ms mean, 33.44 ms median, 37.50 ms at p95. 2B: 33.28 ms, 33.15 ms, 33.55 ms. 4B: 44.16 ms, 44.03 ms, 44.60 ms. All measured per query on a single RTX 4090 through the local Hugging Face path.

• Seven-suite average — 0.8B: 79.38. 2B: 84.68. 4B: 90.02.

• Calibration — 0.8B: Brier 0.530, ECE 0.066. 2B: Brier 0.437, ECE 0.100. 4B: Brier 0.347, ECE 0.065.

• Footprint on disk at bf16 — 0.8B: 1.71 GB. 2B: 4.43 GB. 4B: 9.08 GB.

The 2B is quicker on the mean, dramatically tighter at the tail, and more accurate, all at once. So "smaller means faster" is false across this family — twice over, since the 4B is slower than both. The only axis on which the 0.8B wins outright is the one nobody benchmarks: 1.71 GB against 2B's 4.43 GB and 4B's 9.08 GB. If you are placing a scorer in a fixed-memory budget — a small always-on box, a co-tenant GPU, an edge node with a language model already occupying most of the card — that is the entire case, and it is a real one.

The rest of the table is a study in where small models break unevenly. The 0.8B's Typed Decision score is 77.35 against the 4B's 80.55 — three points apart at a fifth of the parameters count. But Jevbench-Original drops from 98.61 to 80.56 and WildJailBreak collapses from 89.86 to 64.48. Whatever those two suites are measuring — the card does not say, and the card gives no suite definitions at all — they are the ones that punish the small checkpoint hardest. A model that holds up on one axis and falls off a cliff on two others is not a uniformly weaker model. It is a model with a specific competence boundary, and you would want to know where your workload sits before committing the fleet to it.

One more caution on the calibration row. The 0.8B's ECE of 0.066 is its best number — better than Jev's 0.095 on the same suite — while its Brier score of 0.530 is worse than Jev's 0.358. Calibrated error and mean squared probability error are not the same measurement, and a table that shows one looking strong and the other looking weak is telling you the distribution is well-shaped near its confidence peaks and fatter than it should be in between. If your system thresholds on confidence, that distinction matters more than the average.

What 1.71 GB actually buys

The inference path in this model is a scorer, not a generator, and the difference in cost is structural rather than incremental. You hand it a shared state, a schema of named questions, and optionally up to eight images. Each question is one of three types: choice (one of an ordered option set), score (an ordinal scale, returned as a probability-weighted expected value), or noul (binary no/yes). The state, schema and a complete assistant JSON skeleton are rendered with one placeholder per field, the model runs a single causal forward pass, and the logits are read at the position immediately before each placeholder. A softmax is taken over only that field's permitted candidate symbols, the checkpoint's fitted calibration is applied, symbols map back to your option values.

Three consequences fall out of that. There is no output token and therefore no output price — a million decisions cost nothing in tokens. There is no decoding loop and no brace to forget, so malformed JSON is not a failure mode. And because each field's answer space is assembled per request, a schema you invent this afternoon needs no retraining: add a question and its options become candidate symbols for that field.

The limits are stated just as plainly. One to sixteen questions, up to 62 options each, up to eight images, and a default 8,192-token cap — with inputs over it rejected rather than truncated. That last clause is a design decision worth dwelling on, because silent truncation is how a classifier quietly starts answering a different question than the one you asked. InternLM fails loudly instead, which is correct and also an operational trap: a long ticket thread plus a schema plus images will blow past 8,192 sooner than you expect, and there is no graceful path when it does.

And the 1.71 GB buys run-time economics you can multiply out. At 33.98 ms per decision, a million decisions is 9.44 hours of one RTX 4090. The 4B needs 12.3 hours for the same million, and gives up ten and a half points of accuracy in exchange for the 7.37 GB you did not have. Neither number includes the cost of the box, and neither includes the part people forget: you are operating a service, and there is no server in the repository.

Qwen 3.8 is not one model, and the cheap end is the one to notice

Qwen 3.8, taken as a standalone name, is Qwen3.8-2.4T-A95B: the 2.4-trillion-parameter sparse mixture-of-experts core Alibaba published as open weights on August 12, 2026, with roughly 95 billion parameters active per token and a custom licence rather than Apache 2.0. The hosted delivery of that same core is Qwen3.8-Max, generally available on a priced API since August 3 and refreshed as a dated September 2 snapshot. They are one brain sold two ways, and the second delivery is the one most people will actually call.

On independent numbers, Artificial Analysis measures the September snapshot of Qwen3.8-Max at an Intelligence Index of 45.4 — fifteenth of 145 models indexed — with an AA Coding score of 76.2 at ninth of 138, GPQA Diamond at 92.8, Humanity's Last Exam at 43.1 and a long-context recall score of 80.3. The open checkpoint trails the API it was distilled from at 39.9. On our own seven-day playground window, Qwen3.8-Max sits at a p50 of 2,578 ms and a p95 of 9,539 ms at 55.5 output tokens per second, with a 1.57% error rate. Qwen3.8-Max is callable on OrcaRouter at provider list price with 0% markup added, so an Alibaba price change is live on our side the same day rather than at the next billing cycle.

The dense sibling is the one to weigh against a 1.71 GB local checkpoint, though, because it is the cheapest way to buy general capability in this family. Qwen3.8-27B is Apache 2.0, carries a native 262,144-token context, and Qwen3.8-27B sits at $0.33 and $2.40 per million tokens on our catalogue. Its Artificial Analysis Intelligence Index is 33.7. It is roughly thirty-two times the parameter count of Intern-Decision-0.8B, still generative, and still the wrong instrument for a closed-set decision at volume — but it is the honest middle option, and the only member of this comparison that is genuinely cheap and genuinely general at once.

Scorer against generator, stated plainly

• Context — Qwen3.8-Max carries a 1,000,000-token window. Intern-Decision-0.8B rejects anything past 8,192. A factor of 122, enforced rather than truncated.

• Input — Qwen3.8-Max takes text, image and video. Intern-Decision-0.8B takes text and up to eight images, and the image tokens count against the same 8,192 budget.

• Output — Qwen3.8-Max writes, reasons, calls tools and emits JSON on request. Intern-Decision-0.8B cannot produce a paragraph, a rationale or a plan. It can only score the options you already thought to list.

• Per-call cost — a realistic triage call of 800 input tokens and 150 output tokens costs about $0.0025 on Qwen3.8-Max, before any reasoning tokens, and its output side is optimistically small because it is a reasoning model. A million such calls is roughly $2,500. Intern-Decision-0.8B has no token price at all: a million decisions is 9.44 hours of a 4090 you own or rent.

• Latency — 2,578 ms at the median on Qwen3.8-Max over the last seven days, inclusive of network and queueing. Intern-Decision-0.8B's 33.98 ms is a bare forward pass on a local card and is not the same measurement; the honest comparison is one order of magnitude, not eighty times.

• Evidence — every Intern-Decision-0.8B figure here is vendor-reported on InternLM's own harnesses and unreproduced by anyone, including the latency. Every Qwen 3.8 figure is either Alibaba's own or an Artificial Analysis measurement, and they are labelled separately above.

• Independent score — Qwen3.8-Max has one. Intern-Decision-0.8B has none of any kind, and no inference provider deploys it.

Two ways to run this pipeline

The operational fork is sharper than the benchmark fork. Intern-Decision-0.8B is a module, not a service. There is no OpenAI-compatible endpoint, no batching server, no health check, no retry policy, and no second replica unless you build one. It loads in one process and answers one dict at a time, and if you need failover you are the failover. That is a fair description of a 853-million-parameter research checkpoint published without a launch note, and it should be priced into the decision accordingly.

The generative half of the same pipeline is a network call, and a network call is what a router is for. Both Qwen3.8-Max and Qwen3.8-27B are reachable behind one OpenAI-compatible key, and automatic failover means a degraded upstream does not become your incident — worth naming here because Qwen3.8-Max's live seven-day error rate on our side is 1.57%, which is low until it is your request. The pattern that makes sense is the one this pairing has been quietly describing: run the cheap closed-set scorer locally because it is fast and costs nothing per call, and route the reasoning step because that is where the price cuts, the model churn and the outages actually happen.

Two things we will not claim. Intern-Decision-0.8B is not one of our routes and will not be until InternLM hosts it somewhere — we are not going to imply otherwise. And we host no InternLM model at all today, so nothing in this article is a pitch to run the subject through us.

What would change the answer

Three open questions, all answerable within days. Does InternLM publish the GitHub repository its own card links to, and with it a description of how these weights were tuned? Does a launch post follow the checkpoints, converting a quiet push into a documented release with a stated deployment story? And does anyone independent reproduce the 79.38 average or the 0.066 calibration error on hardware that is not InternLM's?

Until one of those lands, the honest reading of Intern-Decision-0.8B is narrow and specific: it is the cheapest closed-set scorer in a family of three, on a footprint that fits where its own faster and more accurate sibling does not, published without an announcement and without the one artefact that would tell you what it was tuned for. If your decision is closed-set, your answers are enumerable in a prompt, and 1.71 GB is the number your deployment actually turns on, it is the right pick and there is nothing else like it at that size. If your answer is not enumerable in advance, or your state runs past 8,192 tokens, or you need a rationale you can show a human, then the comparison was never close — and the model you want was never the 853-million-parameter one.