LFM3-1.7B-Base vs MiniCPM5 2B: Two Vendors, Two Scoreboards, One Honest Answer

Liquid AI created the LFM3-1.7B-Base repository on Hugging Face on October 26, 2026 and published the LFM3 family announcement the next day — and that announcement's benchmark table does something unusually useful: it puts the post-trained LFM3-1.7B directly against MiniCPM5-2B, the 2.5B model OpenBMB has been calling 2B-class open-source state of the art since its weights landed in September. On Liquid's own 27-test suite, LFM3-1.7B wins 23 rows and MiniCPM5-2B wins four. The checkpoint in this article's title is not the one in that table, though, and it carries no benchmark numbers of any kind. That gap is the story.

Three things get blurred in every small-model comparison, so it is worth separating them up front: what Liquid measured, what OpenBMB measured, and what neither vendor has published about the base checkpoints that people actually fine-tune. Only the third category describes a fair fight between the two names on this page.

The table Liquid published, and the model it describes

The LFM3-1.7B model card carries 27 rows spread across seven categories — agents, tool use, agentic coding, coding, instruction following, knowledge and long context — with columns for LFM2.5-2.6B, MiniCPM5 2B, Gemma 4 E4B, Qwen3.5 4B and Qwen3.5 9B. Liquid's own model wins all but four of them, and the margins on the agentic rows are not close: LiveClaw 45.60 against 30.37, WebArena-Verified 38.42 against 9.11, AutomationBench 41.29 against 4.71, BrowseComp-Plus 37.00 against 16.83. The same pattern holds in tool use (BFCL V4 71.39 versus 62.65, ToolSandbox 80.36 versus 68.16) and long context, where RULER 128K reads 65.56 against MiniCPM5-2B's 37.96.

MiniCPM5-2B's four wins are all knowledge or multi-turn robustness: τ³-bench at 69.44 against 68.01, GPQA Diamond at 65.15 against 63.37, Humanity's Last Exam without tools at 8.20 against 7.17, and AA-Omniscience accuracy at 9.13 against 7.07. Those are narrow edges on benchmarks where a fraction of a point is often noise. The agentic gaps are not.

Every number in that paragraph is vendor-reported. Liquid ran its own harnesses, chose its own comparison set and published no per-test methodology on the card. As of October 28, 2026, we could find no LFM3 entry on Artificial Analysis, so there is no independent reproduction to weigh against it. Treat the 23–4 split as Liquid's argument, not as an industry result.

What the base checkpoint actually is

Strip away the post-training and you get a 1.73B text-only model with 24 layers: 13 gated short convolutions, five sliding-window attention layers with a 256-token window, and six global-attention layers. The 256K context figure is the real thing — max_position_embeddings in the published config reads 262,144 — not the 256-token window that early coverage confused it with. The vocabulary is 128,000 tokens, the card lists 16 languages, and the licence is the LFM Open License v1.0, whose commercial grant stops applying to a legal entity with annual revenue of $10 million or more.

The training recipe is documented in more detail than most base releases: 22 trillion tokens in total, with the final 3.5 trillion spent on curated mid-training data that includes agentic traces and tool calls, plus 128K and 256K context extensions. At 4-bit precision the checkpoint is described as taking about 1 GB, which is the number that matters if the target is a phone or a laptop rather than a server.

What the card does not give you is any evaluation of the base model itself. Liquid's published scores belong to LFM3-1.7B after SFT, mixed RL, multi-domain on-policy distillation and agentic RL. There is no way, from public material, to separate how much of that 49.37 average came from the 22 trillion pre-training tokens and how much came from the post-training stack. That is not a criticism specific to Liquid — almost nobody publishes base-model evals — but it does mean the name in this article's title has no scoreboard of its own.

OpenBMB's counter-scoreboard

MiniCPM5-2B's card makes a different claim on a different set of axes. Its capability radar scores nine dimensions and reports an average of 53.9, against 51.1 for Qwen3.5-4B, 42.7 for granite-4.2-3B and 33.2 for LFM2.5-2.6B. On that board, the model is 2B-class state of the art and competitive with 4B-class models. It does not include LFM3-1.7B, which did not exist when the card was written.

The two scoreboards cannot be subtracted from each other. They use different tests, different normalisation and different comparison sets — and the one overlapping benchmark, τ³-bench, is exactly where OpenBMB's model edges ahead in Liquid's table. If you put 53.9 next to 49.37 and conclude that MiniCPM5-2B is the stronger model, you have compared a nine-axis vendor average to a 27-test vendor average and learned nothing.

The hardware facts are less ambiguous. MiniCPM5-2B is a stock LlamaForCausalLM with 2,516,756,480 total parameters, 1,981,982,720 of them non-embedding, 42 layers, grouped-query attention at 16 query and 2 KV heads, and a 131,072-token context in the shipped configuration. It is Apache 2.0, it ships as GGUF, MLX, GPTQ and LiteRT builds alongside a draft model for speculative decoding, and OpenBMB released the training corpora with it. That is a finished product with the factory attached.

Base against base is the only clean comparison

Both families ship a pre-trained checkpoint, which makes a like-for-like comparison possible for the first time. MiniCPM5-2B-Base was created on August 27, 2026, is Apache 2.0, and exists for the same reason LFM3-1.7B-Base does: to be post-trained into something else. Neither base card publishes an evaluation table. So the decision has to be made on specification, licence and ecosystem rather than on a score.

That reframing matters because the most common mistake with this pair is to download the Liquid base, prompt it like a chat model, and conclude it is worse than MiniCPM5-2B. It will be. LFM3-1.7B-Base is mid-training output with a chat template and a thinking mode, not an assistant; Liquid's own card says to use LFM3-1.7B for chat, tool calling and agents. OpenBMB's finished checkpoint is the right thing to compare against LFM3-1.7B, and that comparison already exists in Liquid's table. The base-to-base comparison is for people who intend to do the training themselves.

The spec sheet, dimension by dimension

• Parameters — LFM3-1.7B-Base: 1.73B across 24 layers. MiniCPM5-2B: 2.52B total, 1.98B non-embedding, across 42 layers.

• Context — LFM3-1.7B-Base: 256K tokens, input plus output. MiniCPM5-2B: 131,072 tokens in the shipped config.

• Architecture — LFM3-1.7B-Base: tri-hybrid with gated short convolutions plus sliding-window and global attention, loaded with trust_remote_code. MiniCPM5-2B: a standard Llama decoder that runs on unmodified tooling.

• Licence — LFM3-1.7B-Base: LFM Open License v1.0, commercial use conditioned on staying under $10M annual revenue. MiniCPM5-2B: Apache 2.0, no revenue condition.

• Shipped formats — LFM3-1.7B-Base: BF16 weights only on the base card; the post-trained sibling has GGUF, MLX, ONNX, FP8 and NVFP4 builds. MiniCPM5-2B: BF16, GGUF, MLX, GPTQ and LiteRT, plus SFT and mid-training checkpoints and a DSpark draft model.

• Footprint — LFM3-1.7B-Base: about 1 GB at 4-bit, per Liquid. MiniCPM5-2B: a 2.5B dense model with no 4-bit footprint published on the card; expect the usual 2.5B-class memory cost, not 1 GB.

• Languages — LFM3-1.7B-Base: 16 named languages, including Arabic, Hindi, Thai and Vietnamese. MiniCPM5-2B: an English- and Chinese-centred card, with the released data behind it.

• Training data — LFM3-1.7B-Base: 22T tokens, recipe described but data not released. MiniCPM5-2B: UltraData corpora released alongside the weights.

What the repositories tell you about deployment

MiniCPM5-2B is the easier model to put on a device today. The architecture is standard, so llama.cpp, Ollama and LM Studio support is a matter of using the published GGUF rather than waiting for a kernel. Apple Silicon has MLX builds, Android has a LiteRT path, and the DSpark draft model exists specifically to make decoding faster on constrained hardware.

LFM3-1.7B-Base asks more of the runtime. It needs custom modelling code, and the base checkpoint is published in BF16 only. The surrounding family is better served: Liquid lists vLLM, SGLang and Transformers support for BF16, SGLang for FP8 and NVFP4, llama.cpp and Ollama for GGUF, MLX for Apple Silicon and ONNX Runtime for cross-platform work. The catch is that those builds belong to LFM3-1.7B, not to the base checkpoint. The SGLang pull request that adds LFM3 support, #46858, was still open at the time of writing; the companion cookbook pull request has merged.

If the question is which model to put behind an endpoint this week, note that neither MiniCPM5-2B nor LFM3-1.7B-Base is in OrcaRouter's catalogue today. Both are download-and-self-host propositions for now. That is exactly the situation where a routing layer earns its keep later rather than sooner: one OpenAI-compatible key for 200+ models, 0% markup with the provider's list price passed through, and automatic failover when a self-hosted box or a single upstream goes down. You do not need any of that to run a 1 GB model on a laptop; you need it when the small model becomes one arm of a production path and something has to catch the failure.

What is not knowable yet

The honest list is short and specific. We could find no independent benchmark for any LFM3 model, so every capability claim about the family is Liquid's own. There is no published evaluation of LFM3-1.7B-Base, so nobody can say how good the raw checkpoint is before post-training. There is no per-token price for LFM3 on a hosted API, which makes cost comparisons with MiniCPM5-2B — also self-hosted in most deployments — a hardware-and-electricity exercise rather than a rate-card one. And there is no evidence yet, independent or otherwise, on whether the 256-token sliding window costs the model anything on retrieval-heavy long-context work; Liquid's own RULER 128K score of 65.56 is well below the 87.11 it reports for Qwen3.5 9B, which is a signal worth watching rather than a verdict.

Three questions worth answering before you download either

Do you want a model, or a starting point?

If you want something that answers questions, calls tools and survives a demo on Monday, MiniCPM5-2B is the finished article and LFM3-1.7B-Base is not. If you have a domain corpus, a post-training pipeline and a reason the stock model will not do, the base checkpoint is the correct purchase — and Liquid's TRL notebooks for SFT, DPO, GRPO and continued pre-training are more guidance than most vendors provide.

Will your organisation ever clear $10 million in annual revenue?

The LFM Open License v1.0 grants commercial rights only below that threshold. Above it, the work is not licensed under the agreement at all, and you need a separate arrangement with Liquid. Apache 2.0 has no such cliff. For a hobby project or an early startup the difference is theoretical; for an established company it decides the shortlist before any benchmark is run.

Is 256K context worth a custom architecture?

LFM3-1.7B-Base offers twice the context of MiniCPM5-2B in a smaller parameter count, which is the strongest technical argument in its favour. The price is a non-standard model that needs remote code, a narrower set of ready-made builds on the base checkpoint, and a runtime integration that is still landing. If your workload fits in 131K, you are paying that price for headroom you will not use.

Which one to pick

For most teams comparing these two names, the answer is MiniCPM5-2B — not because it wins a benchmark contest, but because it is a completed Apache-2.0 product with quantised builds, released training data and no licence ceiling. It is the lower-risk way to get a capable 2.5B model onto a device.

LFM3-1.7B-Base is the more interesting bet for a narrower audience: teams that want a 1.73B checkpoint with 256K context, 16 languages and a non-standard architecture they can continue training on a single GPU, and that can live with the LFM licence and a thinner base-model ecosystem. Liquid has given that audience a real starting point and an unusually detailed training recipe. It has not yet given anyone a reason to believe the base checkpoint is better than OpenBMB's finished one, and the vendor's own table does not try to claim that it is.

The number to watch is not the 23–4 split. It is the first independent evaluation of any LFM3 model — the point at which Liquid's scoreboard stops being the only scoreboard.