Shanghai AI Lab released Intern-Decision on 26 September 2026 as Apache-2.0 weights fine-tuned from Qwen3.5, with a paper following on 15 October. It suits agent builders who want a self-hosted router or classifier that decides in tens of milliseconds on one RTX 4090, per the authors' own test.
What changed
Intern-Decision is an open-weights family of 0.8B, 2B and 4B multimodal decision models from Shanghai AI Lab that return typed choices with probabilities in one forward pass. Its authors report 90.02% average accuracy for the 4B model across seven suites, against 88.74% for Jev v1.13.0.
Provider: Shanghai AI Lab · Family: Intern-Decision
More about Shanghai AI Lab on HokAI
Input modalities: text, image · Output: text
About Intern-Decision
Intern-Decision is a set of three small decision models, sized 0.8B, 2B and 4B, built by Shanghai AI Lab. The lab published the training, inference, evaluation and demo code on GitHub and, via the Hugging Face hub, the checkpoints, all on 26 September 2026, and followed with a paper on 15 October 2026. Each model reads a shared state, optional images and a list of typed questions, then returns a choice or score with candidate probabilities instead of a written answer. The three question types are choice, score and yes/no. The backbone is Qwen3.5 from Alibaba: the language model is fine-tuned while the vision tower and projector stay frozen, and the lab keeps the upstream Qwen notices with the weights.
The training objective is what makes it fast. The lab calls it autoregressive masked language modeling: the assistant message is a complete JSON skeleton with one decision placeholder per field, and the answer symbol only ever appears in the training labels. At inference every field is scored in a single causal forward pass, a softmax runs over only the legal single-token symbols, and a per-checkpoint temperature (about 1.99 for the 4B model) adjusts how spread out the probabilities are. Nothing is sampled and generate() is never called. That also means the output is always inside your schema, though the authors are clear that a valid answer is not automatically a correct one, and the fields are not guaranteed to be consistent with each other.
On accuracy, the authors report that the 4B model averages 90.02% over seven suites, ahead of the 88.74% they list for Jev v1.13.0, the hosted model sold by TypeSafe AI. The average is the unweighted mean of Jevbench Easy, Original and Hard, Typed Decision, ToolACE, AG News and WildJailBreak, and the lead is uneven. The 4B model scores 73.87% on Jevbench Hard to Jev's 72.07%, 80.55% on Typed Decision to Jev's 73.35%, and 96.45% on ToolACE to Jev's 91.29%, but Jev leads on WildJailBreak, 96.29% to 89.86%. The 2B model averages 84.68% and the 0.8B model 79.38%. Treat the table as vendor-reported: the repo says protocols and scope differ between systems, part of the public Jevbench set informed development, and the paper has no matched-data ablation to separate model design from training data.
Speed and calibration are the other two claims. On one RTX 4090 in BF16 with native Hugging Face inference, a 289-token request with three fields averaged 44.16 ms for the 4B model and about 34 ms for the smaller two, while the authors list Jev at 109.70 ms under serving conditions they say are incompletely documented, so it is not a controlled hardware comparison. On a separate 96-case pilot with exact reference distributions, temperature scaling cut the 4B model's multiclass Brier score from 0.628 to 0.550 and its expected calibration error from 0.213 to 0.089, against 0.595 and 0.130 for Jev. In a handoff study on 111 Jevbench Hard items, sending the least confident items to Atria Dawn Preview lifted accuracy from 73.87% to 87.39%.
The weights are free to download and the lab runs no hosted API for them that we could find, so you pay only for your own GPU time. That is the main contrast with Jev and Cloudflare Clef from Cloudflare, which are hosted, and the closest open counterpart on HokAI is Strands Decider 2B. Choose Intern-Decision if you want a self-hosted router, classifier or game-and-browser controller that can look at a screenshot; skip it if you need a managed endpoint with an uptime commitment, an option menu longer than 62 labels per field, or a model that writes prose. The Jev alternatives guide compares the whole group.
Pricing
Free to download under Apache-2.0 per the Hugging Face model card; the lab runs no hosted API for the checkpoints that we could find on 2026-10-18. You pay only for your own GPU or cloud time. Hosted competitors meter input tokens, while [Strands Decider 2B](/hub/models/strands-decider-2b) is the other open-weights option.
Key Features
- Several decisions, one forward pass: A request can carry many typed questions, and the answers for all of them come out of one pass through the backbone language-model head.
- Typed answers with probabilities: Choice, score and yes/no questions return the winning label plus candidate probabilities, with a per-checkpoint temperature applied.
- Optional image input: Built on a Qwen3.5 multimodal backbone with the vision tower and projector frozen, so a screenshot can sit in the state.
- Three sizes: The smallest, middle and largest checkpoints trade accuracy for speed, and the authors time all three between about 33 and 44 ms on one RTX 4090.
- Full training and evaluation code: The repo ships training, two inference backends, the seven accuracy suites, temperature presets and a 96-case calibration benchmark.
Pros
- Open weights and code under Apache-2.0 remove the per-call bill and keep data on your own machine.
- Image input makes it usable for screen and game controllers, where text-only decision models need a separate captioner.
- Ships its own calibration benchmark and temperature presets, so confidence scores are meant to be used, not ignored.
Cons
- Headline numbers are vendor-reported on suites that partly informed development, so they are not independent proof.
- No hosted endpoint or SLA from the lab; you operate the GPU and the serving code.
- Output is typed choices only, so it cannot replace a general chat model.
Frequently Asked Questions
How much does Intern-Decision cost in 2026?
The weights cost nothing: the Hugging Face model card lists Apache-2.0, and we found no hosted API from the lab on 18 October 2026. Your bill is your own GPU or cloud time, and one consumer RTX 4090 was enough for the authors' latency test. Hosted rivals such as Jev charge per input token, so the choice is hardware you run against a metered endpoint.
Does Intern-Decision have a free plan?
There is no plan to buy, only files to download. The code and checkpoints are public, and the lab hosts a browser demo on Hugging Face Spaces for trying the interface. The demo examples are synthetic illustrations, not production traffic.
Which tools compete with Intern-Decision in 2026?
Jev is the proprietary, hosted benchmark it was built to beat, and Cloudflare Clef is another hosted decision model. Strands Decider 2B is the closest open option, also Apache-2.0 and self-hosted. Pick a hosted one if you want someone else to run the servers.
What separates Intern-Decision from Jev?
Intern-Decision gives you weights, image input and a published recipe; Jev gives you a managed API and, in the authors' own table, a lead on the WildJailBreak suite. The paper says the speed comparison is not a controlled hardware test, so treat both the accuracy lead and the speed gap as the lab's claims. Take Intern-Decision if you must self-host or read screenshots, Jev if you want an endpoint.
How do you get started with Intern-Decision?
Use Python 3.12 or newer, clone the GitHub repo, install its requirements and download a checkpoint from the Hugging Face collection. Then build a request with a state, ordered typed questions and optional images, keep option order fixed, and keep the checkpoint default temperature. The repo includes a demo and evaluation scripts to confirm your setup.