Intern-Decision-0.8B vs LFM2.5 2.6B Base: Two Ways to Build Something Yourself

Put Intern-Decision-0.8B next to LFM2.5 2.6B Base and the honest first observation is that neither one is a product in the sense a buyer usually means. Intern-Decision-0.8B is the checkpoint InternLM uploaded to Hugging Face on 26 September 2026 with no announcement at all — a 852,985,920-parameter fine-tune of Qwen3.5-0.8B that answers typed questions and emits no text, shipped under Apache 2.0 with a GitHub link that 404s. LFM2.5 2.6B Base is Liquid AI's 2.69-billion-parameter pre-trained text checkpoint, posted on 1 August 2026 as the foundation for the LFM2.5 family, and its own card says it is "only recommended for tasks that require heavy fine-tuning." One is finished and narrow. The other is unfinished and general. Both assume you are the one doing the work — and the work is not the same work.

That distinction is the whole comparison, and it is the reason a straight spec table would mislead. The question a reader actually has is "which of these do I download," and the answer depends on whether the thing you need to build is a decision layer or a language capability. Those are different projects with different timelines.

What each model hands you

Intern-Decision-0.8B arrives as a working system. The repository ships inference.py/, a DecisionEngine/ class, a documented request schema and a response schema, and a worked example you can run after installing four pinned Python dependencies. You supply a state, one to sixteen named questions with up to 62 options each, and optionally up to eight images, and you get back a calibrated distribution plus an argmax label per field. The engine reads logits at fixed positions in a pre-rendered skeleton rather than generating anything, so there is no decoding step, no output tokens, and no JSON to repair. It runs from roughly 1.73 GB of files.

LFM2.5 2.6B Base hands you the opposite: raw capability with no interface. It is a text-only causal language model with 30 layers arranged as 22 double-gated short-convolution blocks and 8 grouped-query-attention layers, pre-trained on approximately 34 trillion tokens across 16 languages, with a 128,000-token vocabulary and a 131,072-token context window. It has no instruction tuning of its own and no task benchmarks on the card, because a base checkpoint is not scored — the models you train from it are. Liquid's own post-training pipeline is what turns it into the agentic LFM2.5-2.6B, and that pipeline is the thing you are being asked to reproduce, partially or wholly, for your own domain.

So the axis is not size. It is whether the missing piece is a decision contract or a training run.

The scoreboard, with the caveats attached

• Time to first result — Intern-Decision-0.8B: an afternoon, loading a checkpoint and calling predict()/. LFM2.5 2.6B Base: however long your continued pre-training or SFT run takes, plus the data work to prepare it.

• Parameters — Intern-Decision-0.8B: 852,985,920 across a 1.50 GB language shard, a 176 MB vision shard and a 25 MB projector. LFM2.5 2.6B Base: 2.69B, roughly 2.5 GB of weights.

• Input modalities — Intern-Decision-0.8B: text plus up to eight images, with vision weights present in the repo. LFM2.5 2.6B Base: text only, by design — the multimodal work in the LFM2.5 family lives in the VL variants.

• Practical context — Intern-Decision-0.8B: 8,192 tokens, rejected rather than truncated, despite a 262,144 maximum position embedding in the config. LFM2.5 2.6B Base: 131,072 tokens native.

• Output — Intern-Decision-0.8B: a calibrated probability distribution and an argmax label per field, deterministic. LFM2.5 2.6B Base: continued text, sampled.

• Benchmarks — Intern-Decision-0.8B: a full vendor table across seven benchmarks, unreproduced by anyone. LFM2.5 2.6B Base: none published for the base itself, by design.

• Licence — Intern-Decision-0.8B: Apache 2.0, with a preserved LICENSE-QWEN/ for the upstream weights. LFM2.5 2.6B Base: LFM Open License v1.0.

The licence row deserves its own section

Both cards say "open," and the two words mean materially different things. Intern-Decision-0.8B is Apache 2.0 — no revenue condition, no field-of-use restriction, no separate commercial grant to negotiate. The second licence file in the repository is the Qwen licence carried forward because the model is a derivative of Qwen3.5-0.8B, which is the correct handling rather than a limitation.

LFM2.5 2.6B Base ships under the LFM Open License v1.0, and section 5 of that licence is worth reading in full before any commercial plan depends on it. The rights granted for Commercial Use are conditioned on you or your legal entity not exceeding a threshold defined in the licence as annual revenue of 10 million US dollars or more. Above that line, commercial use of the work or a derivative is not licensed under the agreement. Non-profit and research use is carved out. That is a real and enforceable boundary, and because it attaches to derivative works too, it propagates into anything you fine-tune from the base — which is the entire intended use of this checkpoint.

There is a straightforward reading here: if you are building commercially and your entity clears $10M in annual revenue, LFM2.5 2.6B Base is not the same kind of asset as Intern-Decision-0.8B, regardless of how the two perform. If you are under the threshold, researching, or fine-tuning for a non-commercial purpose, the distinction does not bite. Either way it is a question to answer before the training run, not after.

Where each one is actually the right download

LFM2.5 2.6B Base is the right choice when the capability you need does not exist yet in a finished model. Liquid's card lists the intended cases explicitly: language-specific assistants such as Japanese, domain-specific assistants such as medical, training on proprietary data you cannot send to an API, or experimenting with novel post-training approaches. The reasoning behind that list is that 34 trillion tokens of multilingual pre-training is expensive to reproduce and nearly free to inherit — you are buying a starting point that is already competent across 16 languages and a 128K context, and spending your own compute on the part that is specific to you.

The ecosystem support for that plan is unusually complete for a model this new. Liquid documents continuation pre-training, SFT, DPO and GRPO through Unsloth and TRL, with notebooks for each, and the checkpoint is supported by Transformers, vLLM, SGLang, llama.cpp, MLX and LM Studio. That is not marketing copy — it is the difference between a base model you can fine-tune this week and one you spend a fortnight gluing into a trainer.

Intern-Decision-0.8B is the right choice when the capability you need does already exist and the problem is its shape. If your task is a bounded decision with a closed answer set — route this ticket, score this claim, classify this record against a rubric — then a generative model is an awkward way to get a label, and a base model is not a way to get one at all. What you want is something that returns the distribution, tells you its confidence, and does it in the same 34 milliseconds every time. InternLM's measured latency on a single RTX 4090 is 33.98 ms mean with a 37.50 ms p95, and the card's own numbers show the 2B sibling at 33.28 ms mean — a reminder that at these prompt lengths the cost is reading the schema, not running the weights.

The trade you are making is flexibility for a contract. A base model can become anything, eventually, if you have the data and the compute. A decision head is already the thing, and it will never become anything else. If your schema changes shape every month, that is a real cost. If it does not, it is not a cost at all.

What nobody can tell you about the Intern-Decision table

Every performance number for Intern-Decision-0.8B is vendor-reported, and the caveat here is heavier than usual because the release is a week old and completely undocumented. There is no paper, no collection gathering the three sizes, no working GitHub repository, and no independent evaluation. A search of Artificial Analysis returns nothing for the model.

InternLM selected the benchmarks, selected the comparison models — a set dominated by decision models including TypeSafe's Jev, Convai's Laya and Kev — and published the table without a launch post.

With that label attached, the shape is still worth reading. Intern-Decision-0.8B posts 97.92 / 80.56 / 52.25 across the three Jevbench splits, 77.35 on Typed Decision and 94.52 on ToolACE, averaging 79.38. Its calibration profile is the most interesting entry: a Brier of 0.530 and an expected calibration error of 0.066, which is a better ECE than Jev's 0.095 despite a worse Brier. In plain terms it is less accurate but more honest about how accurate it is, and for a threshold-based workflow that ordering matters more than the average. The column that should stop a deployment is WildJailBreak at 64.48 against Jev's 96.29. A refusal-robustness deficit that large is a property of the fine-tune, and whether it belongs to the Qwen3.5-0.8B base, the decision-tuning objective or the parameter count is not documented anywhere on the card.

The comparison against LFM2.5 2.6B Base on this axis is not "who scores higher," because the base has no scores. It is that one model's numbers are unaudited and the other model's numbers do not exist, and a reader choosing between them is choosing which kind of unknown to work with.

The pipeline most people actually end up building

There is a configuration that uses both, and it is worth naming because it is where most production systems land. The decision layer handles the closed calls — triage, routing, rubric scoring — where the answer set is known, latency compounds across a million records, and output tokens are pure overhead. The language layer handles everything that needs prose, synthesis or long context: the escalation summary, the explanation attached to the label, the draft a human edits. LFM2.5 2.6B Base is a plausible substrate for the second half if you have domain data worth training on; a decision head is the natural shape of the first.

Composing the two is the fiddly part, and it is the part worth not building by hand. OrcaRouter's routing DSL expresses routing as code — YAML with CEL conditions, deployed without a redeploy — so a pipeline that sends one class of request to a decision endpoint and another to a language model is a routing rule rather than a load balancer you maintain yourself. The gateway covers 200+ models behind one OpenAI-compatible key with provider list price passed through at zero markup, and it does not add a markup on the model you pick. To be precise about the boundary: neither Intern-Decision-0.8B nor LFM2.5 2.6B Base is one of ours — both are checkpoints you download and run locally, and in both cases that is the point, because the reason to pick a 2 GB model is that it runs where your data already is. What a gateway is for is the half of the stack you would otherwise be calling out to.

If the decision layer is the piece you want to test without committing a training run to it, a hosted decision model is the cheaper experiment, and TypeSafe's Jev 1.13 is the one InternLM benchmarked against — callable today through a single OpenAI-compatible endpoint at $0.042 per million input tokens with output billed at zero.

Which one, then

Choose LFM2.5 2.6B Base if the capability does not exist yet and you have both the domain data and the appetite for a fine-tuning run, and check the $10M revenue threshold in the LFM Open License before you build on it commercially. Choose Intern-Decision-0.8B if the capability exists, the answers are already enumerable in your prompt, and what you need is a fast, deterministic, licence-clean way to get the label — accepting that its documentation is missing, its benchmarks are unaudited, and its refusal behaviour on adversarial input has not been tested by anyone outside the lab that tuned it. The second is the shorter path and the larger unknown. The first is the longer path and the more documented one, and neither of those facts is the same as better.