Intern-Decision-2B vs Qwen 3.8: One Backbone, Two Very Different Jobs

Open the config for internlm/Intern-Decision-2B and the config for Qwen/Qwen3.5-2B side by side and almost nothing differs. Same 24 layers, same 2,048 hidden size, same 8 query heads against 2 key-value heads, same head dimension of 256, same 262,144-position embedding ceiling, same 248,320-token vocabulary, same repeating pattern of three linear-attention layers to one full-attention layer. Intern-Decision-2B is not a new architecture. It is that backbone with its ability to generate text removed and its confidence calibrated — and that single subtraction is what makes the comparison against Qwen3.8-Max, the 2.4-trillion-parameter flagship that the whole "Qwe​n 3.8" family name refers to at the top end, worth more than a parameter countdown.

The two are not competitors and the matchup is not a contest. Intern-Decision-2B is a 2,213,241,664-parameter fine-tune of Qwen3.5-2B, uploaded to Hugging Face at 05:36 UTC on September 26, 2026 among three unannounced siblings, and it emits no tokens: it takes a state and typed questions, runs one causal forward pass, reads logits at fixed placeholder positions, and returns a calibrated distribution per field. Qwen3.8-Max is Alibaba's hosted flagship, a sparse mixture-of-experts model with roughly 95 billion parameters active per token, a 1-million-token multimodal context window and list pricing of $2.00 per million input tokens and $6.00 per million output. One is a component. The other is a service. The useful question is where the seam between them belongs in a pipeline you are already paying for.

The subtraction, precisely

What decision tuning changed is worth stating exactly, because it clarifies what the base model was already carrying.

The task is named in the repository as autoregressive masked language modelling. The assistant input contains a complete JSON skeleton with one <decision> token per field; the ground-truth answer symbols appear only in the labels, never in the input. Training uses ordinary causal next-token alignment, where the logit immediately before a marker predicts that field's answer, and all fields share one forward pass. At inference the logits are restricted to the legal single-token answer symbols — A–Z, a–z, 0–9, which is where the 62-option ceiling comes from — then softmaxed and mapped back to your labels.

So the base model's next-token machinery is intact and unmodified. What was trained in is a preference for occupying one of a handful of answer positions instead of continuing a sentence, plus a checkpoint-specific temperature of 2.100509348278 fitted by negative log-likelihood over 1,728 designated calibration cases with 1,693 held out. The card is unambiguous that generation is off the table: the API "performs structured candidate scoring. It does not call generate() or sample free-form text."

The repository also records what the tuning did not touch: it "fine-tunes the language backbone of Qwen3.5 while freezing the vision tower and projector." The 612,517,440-byte vision shard on the 2B is the same byte count as on the 4B decision checkpoint, which fits inherited rather than trained components. The image path works; it simply was not what the decision pass spent its budget on.

What 2.2 billion parameters cannot do, and what the 2.4 trillion do instead

Everything separates on one axis: whether your answer already exists as a list.

Intern-Decision-2B answers a closed set. One to sixteen questions, up to 62 options each, up to eight images, all inside an 8,192-token budget that the engine refuses rather than truncates. Returned as probabilities. At 33.28 ms mean per request on a single RTX 4090 with a P95 of 33.55 ms, and nothing billed per call because there is no output to bill for.

Qwen3.8-Max writes. A 1-million-token context, text, image and video input, reasoning with a thinking mode, native tool calling, structured outputs, up to 131,000 output tokens on the dated 0902 build. It can also, in principle, make the same routing decision Intern-Decision-2B makes — you can prompt it to choose among options and return JSON. Three things go wrong when you do. You pay for the tokens it spends deciding. You get a string whose parse may or may not succeed. And the number inside that string is whatever the sampler produced, not a softmax you can threshold at 0.8 and act on.

Both accounts are true and neither cancels the other. The 2.4-trillion-parameter flagship exists because most problems are not closed sets. The 2.2-billion scorer exists because the subset that is deserves a cheaper instrument.

Scoreboard

• Parameters — Intern-Decision-2B 2,213,241,664 in BF16 plus a vision tower and projector, about 4.46 GB on disk; Qwen3.8-Max roughly 2.4 trillion total with about 95 billion active per token, served not downloaded.

• Context — 8,192 tokens, rejected above; 1,000,000 tokens with 131,000 maximum output on the 0902 build.

• Output — calibrated probabilities and an argmax, no generated text; generated text including a reasoning trace.

• Modalities in — text plus up to eight images; text, image and video.

• Cost — no per-call charge, and you own the GPU; $2.00 per million input tokens, $6.00 per million output, with cache reads at $0.25 and cache writes at $2.50.

• Published evidence — a seven-suite vendor table averaging 84.68, Brier 0.437 and ECE 0.100 across 10,751 test rows, unreproduced by anyone outside InternLM, on a checkpoint with one like and zero downloads; an Artificial Analysis Intelligence Index of 45.4 at rank 15 of 145 and an AA Coding index of 76.2 at rank 9 of 138, both independently measured.

• Latency — 33.28 ms mean per request on one RTX 4090, going down as batching is added; 2.655 seconds P50 to first token as the catalogue reports it, going up with load.

The evidence rows are not equivalent and should not be printed as though they were. Alibaba's flagship has an independent index score behind it. InternLM's checkpoint has a table its own authors produced, and the calibration figure especially should be read as a well-documented intention until it meets somebody else's data — the more so here, because this particular size posts the worst expected calibration error of the three InternLM shipped, 0.100 against 0.066 for the model half its size and 0.065 for the one nearly twice it.

The seam, and how to run it

The pipeline that makes sense is not a choice between these two but a split: the scorer handles the enumerated decisions, and a frontier model handles everything whose answer is not a list. A support triage that routes to one of six teams and rates urgency on a three-point scale does not need 95 billion active parameters; it needs a probability and a threshold. The escalation path that eventually writes the customer a reply does.

That split has an operational form. Intern-Decision-2B is not on OrcaRouter — our model page for it returns 404 and there is no hosted route anywhere we could find — so it runs on your own hardware, which means it is a component you operate and monitor. Qwen3.8-Max is a route: one key across 200-plus models, provider list price passed through with no markup, which means a vendor price change on the flagship reaches your bill the same day it lands. Putting the hosted leg of the pipeline behind that single surface is what keeps the cheaper leg cheap: the scorer keeps the call volume down, and the frontier model is only reached for the cases that actually need it.

If you are still evaluating which scorer belongs in that slot — this one has no independent evaluation, and its calibration is the weakest in its own family — automatic failover across providers is the way to try it in a path that already has a working alternative behind it rather than replacing that alternative outright.

The honest verdict

If your problem is a decision over a set of answers you can enumerate, and the state fits inside eight thousand tokens, Intern-Decision-2B will do it in thirty-three milliseconds for nothing per call, and no frontier model will beat it on that axis no matter how many parameters you rent. Get your own calibration onto it before you depend on the confidence it reports.

If your problem is anything else — reasoning, long context, tool use, video, a million-token document, or simply an answer that has not been written yet — Qwen3.8-Max is not merely better. It is the only one of the two that can do the job at all.

And if you cannot yet say which of those two you have, the answer is probably that you have both, in the same pipeline, which is the architecture the parameter counts were quietly telling you all along.