LFM3-1.7B-Base vs Granite 4.2 3B: Raw Material Against a Finished Reasoner

A base model that ships with a chat template is a contradiction most vendors avoid. LFM3-1.7B-Base, published on October 26, 2026, contains a full chat template and turns reasoning on by default when you apply it — yet its own card says it is only recommended for heavy fine-tuning and directs chat traffic elsewhere. Granite 4.2 3B, IBM's Apache-2.0 reasoning model from August 25, 2026, is the opposite kind of artifact: post-trained, benchmarked, three reasoning modes, tool calling, and a card that claims 78.33 on AIME 2025. The interesting question is not which is better. It is why the half-finished one behaves like a finished one, and what that tells you about which to download. A fallback path should not cost a premium. list prices Provider list prices pass through without markup.

The odd one is the base model

Liquid's decision to mid-train on chat data with and without thinking traces is the detail that makes this comparison more than a size contest. The last 3.5 trillion of the checkpoint's 22 trillion pre-training tokens were curated with agentic traces and tool calls, formatted with the same chat template the tuned model uses. That is why the base repository contains a template at all, and why a naive apply_chat_template call produces a thinking block before an answer. The model has seen the format. It has not been trained to follow instructions in it.

That distinction matters in practice. A fine-tune starting from LFM3-1.7B-Base inherits a model that already recognizes the conversation structure and the reasoning delimiters, which shortens the supervised phase. It does not inherit reliability, refusal behavior or tool-call correctness. Those come from the post-training you are about to do, which is exactly what the card says the checkpoint is for.

What Granite 4.2 3B actually is

IBM's 3B is the edge member of the Granite 4.2 reasoning family, post-trained from Granite-4.1-3B-Base. It is dense, text-only, and runs at 128K native context with an extension to 512K. Three reasoning modes ship with it — full thinking by default, a low-effort mode, and off — plus OpenAI-compatible tool calling and a family of quantization formats. It runs in six to eight gigabytes in bfloat16, under two gigabytes at a practical quantization, and Ollama lists a granite4.2:3b build.

The card reports 78.33 on AIME 2025 in full thinking mode, 54.80 on GPQA, and 69.71 on LiveCodeBench v6. Every one of those is IBM's own number, printed the day the weights went up, and none has been independently reproduced. The 3B also lacks the agentic reinforcement learning that its 8B and 30B siblings received, which is why it carries no SWE-Bench or Terminal-Bench scores. It is a reasoning model, not an agent engine.

What the LFM3 checkpoint actually is

A 1.73-billion-parameter pre-trained text model with a 262,144-token context, a 128,000-token vocabulary across sixteen languages, and 3.46 GB of bfloat16 weights. Twenty-four layers: thirteen gated short convolutions, five sliding-window attention blocks, six full-attention blocks. About 1 GB at 4-bit. The LFM Open License v1.0, free for commercial use below $10 million in annual revenue. Two TRL notebooks on the card, one for supervised fine-tuning on structured extraction and one for continued pre-training on a new language. No benchmark table of any kind.

The five numbers that separate them

• Parameters — 1.73B dense for LFM3-1.7B-Base against roughly 3B for Granite 4.2 3B. Neither is a mixture of experts; every parameter is read on every token.

• Pre-training tokens — 22 trillion for LFM3-1.7B-Base. IBM says the Granite 4.2 family saw roughly 15 trillion across five phases, for a model nearly twice the size. The smaller checkpoint saw more text.

• Context — 262,144 tokens native for LFM3-1.7B-Base; 128K native and 512K extended for Granite 4.2 3B. The base model wins on native length, the tuned model wins on the ceiling.

• Benchmarks — none published for LFM3-1.7B-Base. AIME 2025 78.33, GPQA 54.80 and LiveCodeBench v6 69.71 for Granite 4.2 3B, all vendor-reported and unreproduced.

• License — LFM Open License v1.0 with a $10 million revenue threshold against Apache 2.0 with no commercial trigger.

Why the template exists, and what it is worth

There is a real engineering benefit hiding in the oddity. Continued pre-training and supervised fine-tuning both need the model to recognize the format it will be served in. Starting from a checkpoint that has already seen chat markup and thinking delimiters means the supervised phase spends its gradient budget on behavior rather than on formatting. Granite 4.2 3B does not give you that starting point; it gives you the finished behavior, which you can then specialize, but the specialization starts from a model that already has opinions about how to answer.

That is the trade in one sentence. LFM3-1.7B-Base gives you a cleaner slate with a familiar shape. Granite 4.2 3B gives you a working reasoner you can point at a problem today and, if you want, fine-tune afterward — with the caveat that fine-tuning a post-trained reasoning model can erode the reasoning behavior you paid for unless the data is built carefully.

The budget question nobody puts in the table

Inference is the cheap side of this comparison. Both models run on a laptop, both quantize below 2 GB, and at that size the difference between 1.73B and 3B is a matter of tokens per second rather than feasibility. The expensive side is data. Fine-tuning either model on a proprietary task requires labelled examples, an evaluation set and a training loop, and that cost dwarfs the difference between a 3.46 GB download and a 6 GB one.

What changes with size is how much of that budget goes into capability the base model already has. A 1.73B trunk trained on 22 trillion tokens is a deliberate bet that a small model with a long context and a broad language mix can absorb a domain quickly. A 3B reasoning model is a bet that general reasoning transfers to your domain with less data. Neither bet is settled by the vendor's numbers, because neither vendor published a fine-tuning result on a task you care about.

Once a fine-tune exists, the serving question is separate. OrcaRouter's routing DSL lets a specialist and a general fallback sit behind one call, and automatic failover means a small model timing out does not take the request with it. That matters more for a 1.73B fine-tune than for a large model, because the failure mode of a small specialist is confidence in the wrong domain, and a fallback is the cheapest guard against it.

License: the clause that decides enterprise adoption

Apache 2.0 is the shortest license conversation in procurement. Granite 4.2 3B ships under it, with no revenue threshold and no commercial trigger. LFM3-1.7B-Base ships under the LFM Open License v1.0, which grants free commercial use below $10 million in annual revenue and requires a commercial license above it. For an independent developer the difference is invisible. For a product team inside a larger company, the threshold is a legal review that happens before the first training run, and it can outweigh a parameter-count advantage in either direction.

Which one to download

You need a reasoning model this week. Granite 4.2 3B. It is finished, documented, quantized and small enough for a laptop. The vendor numbers are unreproduced, but the model runs and the interface is standard.

You are building a domain specialist and have the data. LFM3-1.7B-Base. The 22-trillion-token trunk, 256K context and sixteen-language vocabulary are a better starting point than a reasoning model you would partly overwrite, and the chat template means the supervised phase starts from a familiar format.

You need both a reasoner and a specialist. They are not substitutes. Granite 4.2 3B handles the general reasoning path; an LFM3 fine-tune handles the narrow, high-volume path. Running both behind one endpoint is the pattern that avoids choosing.

What is still missing

The gap in this comparison is not a number, it is a reproduction. Granite 4.2 3B has vendor benchmarks and no independent runs; LFM3-1.7B-Base has neither. Until someone fine-tunes the Liquid checkpoint on a public task and reports the result against a Granite 4.2 3B baseline, the choice rests on what kind of asset you want rather than which model performs better. The template makes the base checkpoint look more finished than it is. The card is honest about that, and it is the part worth reading before the download.