Intern-Decision-4B vs ContextPilot-8B: A Model That Shrinks the Context Against a Model That Manages It

Pick your poison: internlm/Intern-Decision-4B refuses to read more than 8,192 tokens, and tencent/ContextPilot-8B exists specifically to survive contexts that grow without bound. They are the same size class, both fine-tunes of a Qwe​n base, and both landed on the Hub with no vendor announcement and no independent evaluation of any kind — ContextPilot-8B on 27 August 2026, Intern-Decision-4B on 26 September 2026 — and they solve opposite problems. Intern-Decision-4B is a 4.5-billion-parameter structured decision model that runs one forward pass, reads logits, and returns a calibrated probability for every question you asked; it never writes a token. ContextPilot-8B is a 8.19-billion-parameter text generator trained to plan, remember and offload its own working context during a long agent run. Comparing them is not really a benchmark question. It is a question about which half of your problem you would rather have the model swallow.

First, the thing they genuinely share

Neither model has a single number that anyone outside its own lab has verified. That is unusual enough to state before anything else, because most comparisons on this blog can at least lean on an independent leaderboard for one side.

Intern-Decision-4B publishes a full evaluation table on its card — a 90.02 average across seven sets, a Brier score of 0.347 and an expected calibration error of 0.065 — all of it measured by InternLM on InternLM's harness. ContextPilot-8B publishes no table at all. Its card says the framework "consistently outperform[s] existing baselines across various base models and benchmarks" and points at a paper and an evaluation pipeline for the detail. So for this matchup the honest sourcing line is short: one side is vendor-reported and unreproduced, the other is vendor-claimed and unpublished, and nothing on this page is independent.

The two bets, stated plainly

Intern-Decision-4B's bet is that the hard part of a decision is not reasoning, it is committing. Give it a state, a schema of named questions and up to eight images, and it returns a distribution over your own option strings. The inference procedure is deterministic and shape-fixed: map options onto single-token symbols, render an assistant JSON skeleton with one placeholder per field, run one causal forward pass, read the logits immediately before each placeholder, softmax over only that field's candidates, apply the fitted temperature. There is no decoding loop, so there is no parser and no free-text hallucination surface. The price of that bet is a hard input ceiling — the card says inputs above DecisionEngine(max_length=8192) are "rejected without truncation".

ContextPilot-8B's bet is the mirror image: keep the agent writing tokens, but teach it to edit what it is carrying. It adds planning, structured long-term memory, retrieval and soft context offloading to the agent's toolset, then trains the checkpoint with an RL recipe that samples branches around the context-editing decisions that matter and assigns credit at the action level rather than the trajectory level. The card is candid about the packaging consequence: loading the checkpoint does not give you context management. The tool definitions, the agent runtime and the evaluation pipeline live elsewhere and have to be brought along.

Scoreboard, dimension by dimension

• Output — Intern-Decision-4B emits no text at all, only probabilities over your option values; ContextPilot-8B generates normally and its answers are prose and tool calls you have to parse.

• Input ceiling — Intern-Decision-4B refuses anything past 8,192 tokens; ContextPilot-8B inherits Qwen3-8B's 40,960-token position limit and is built to keep working while the effective context grows.

• Parameters — 4.54B BF16 for Intern-Decision-4B, across a text tower, a vision tower and a projector; 8.19B BF16 for ContextPilot-8B, a plain dense Qwen3 causal language model.

• Latency — Intern-Decision-4B runs 44.16 ms mean and 44.60 ms at P95 on a single RTX 4090, per its own card; ContextPilot-8B publishes no latency figure, and its cost is fundamentally different because it is generating tokens rather than scoring them.

• Images — Intern-Decision-4B takes up to eight, and image tokens count against the 8,192 ceiling; ContextPilot-8B's card describes a text-generation checkpoint with no multimodal path.

• License — Intern-Decision-4B is Apache-2.0 with the upstream Qwen license preserved beside it; ContextPilot-8B's license opens with Section 0, "made available solely for the purpose of scientific research and development. You shall not use it for any other purpose."

• Published evidence — a seven-set table with calibration diagnostics on one side; a paper abstract and a pointer to an evaluation repository on the other.

• Track record — both quiet. Intern-Decision-4B shows one like and zero downloads since 26 September 2026; ContextPilot-8B has 11 likes and roughly a thousand downloads since 27 August 2026, which is more attention but still no third-party evaluation.

Where each one actually breaks

The Intern-Decision-4B failure mode is structural, and it is worth being concrete about it. If the evidence for a decision does not fit in 8,192 tokens, the model does not degrade gracefully — it rejects the request. That is a defensible engineering choice and a terrible fit for exactly the workload ContextPilot-8B was built for. The card's own config makes the tension sharper: the text tower under the wrapper declares a 262,144-token position limit. The backbone can address a quarter of a million tokens and the released wrapper will not accept more than eight thousand.

ContextPilot-8B's failure mode is that it is not a drop-in anything. It is a checkpoint inside a framework, and the framework is where the behaviour lives. Pull the weights without the tool definitions and the agent runtime and you have a Qwen3-8B fine-tune whose distinguishing capability is inert. Add to that Section 0 of the license, and the practical answer for a commercial deployment is that you cannot use it at all as shipped — which also means it is not a candidate route for us, now or soon.

Deploying each, if you were going to

Intern-Decision-4B wants a small GPU and no serving stack worth the name: install PyTorch 2.9.1 and Transformers 5.14.1 under Python 3.12, load the four shards once, keep the engine resident, and score. Because the forward pass is fixed-shape, P50 and P95 sit within six-tenths of a millisecond of each other, so capacity planning is about concurrency rather than tail latency. The awkward part is that it ships as an importable Python class rather than an HTTP service, so getting it in front of a production caller means wrapping it yourself.

ContextPilot-8B wants the opposite treatment: a real serving path, a tool executor, memory stores, and a branch-capable rollout environment if you ever intend to retrain or evaluate it as designed. This is not a model you trial in an afternoon; it is a research framework you adopt.

Does a router help here?

Partly, and the honest version is narrower than usual. OrcaRouter does not list Intern-Decision-4B, ContextPilot-8B, or any comparable decision-scoring checkpoint in its catalogue — our model pages 404 for both, and nothing in this piece should be read as an availability claim. What a unified endpoint does buy you is the ability to put a failed experiment behind a fallback: one key across 200-plus models, provider list price passed through with 0% markup, and automatic failover so that a scorer you are still evaluating never becomes a single point of failure. That is a real benefit for the Intern-Decision side of this comparison, where the model actually runs cheaply and locally. For ContextPilot-8B it is moot, because a research-and-development-only license rules out a commercial route regardless of what any router supports.

Which one, then

If your question has a closed set of answers, arrives with the evidence attached, and needs a probability rather than an explanation, Intern-Decision-4B is the more interesting object — cheap, fixed-shape, calibrated by construction, and honest about the input it will not accept. If your problem is that the evidence will not stop arriving, no decision scorer is going to help you and ContextPilot-8B is aimed at the right target, with the caveat that you are adopting a framework and a research license rather than a model.

What would settle it is a test neither vendor has run: hand both the same short, structured decision with the answer computable from the state, and see whether the scorer's 90.02 average holds up outside its own harness. Until somebody does that, the correct reading of this matchup is that the two models are not competing for the same slot at all.