Beam 1.0 vs LFM2.5 2.6B Base: An Unfinished Substrate and a Finished Giant

A model that has not been taught to answer questions is being compared to a model that has been taught to plan. LFM2.5 2.6B Base is Liquid AI's 2.698-billion-parameter pretrained checkpoint, published on 4 August 2026 under the LFM Open License v1.0 and explicitly described as the base for fine-tuning — it is text-only, it is 30 layers of hybrid convolution and grouped-query attention, and it is not post-trained. Reflection Beam 1.0 is a sparse mixture-of-experts model with 501.3 billion total parameters and 23.4 billion active, post-trained through supervised and reinforcement stages and released as downloadable weights on 16 October 2026 under Apache 2.0, following its 5 October 2026 announcement. The size gap is 185 to one. The gap that actually decides the comparison is not the size — it is that one of these is a raw substrate you point a training run at, and the other is a product you serve.

Both are on Hugging Face. Both are downloadable today. Both expect you to bring your own hardware. That is the entire overlap, and reading them against each other is a useful way to see how differently the open-weights ecosystem labels two things it files under the same tag.

The two releases, described as their publishers describe them

LFM2.5 2.6B Base goes by the fuller name LFM2.5-2.6B-Base. Its card lists 2.698 billion total parameters, 30 layers of which 22 are double-gated short convolution blocks and eight are grouped-query attention, a training budget of 34 trillion tokens, a 128,000-token vocabulary and a 131,072-token context length. It covers 16 languages. It is one of a pair: Liquid AI also publishes LFM2.5-2.6B, post-trained for agentic workloads, and this is the raw checkpoint from which every variant in the family descends. The card's own words for the base model are that it is the pretrained text-only checkpoint used to create all the other variants. Sixteen fine-tuned derivatives already exist under it on the Hub.

Beam 1.0 is the opposite arrangement. Reflection ships 501.3B total parameters with 23.4B active per token across 52 layers — three dense and 49 mixture-of-experts, 256 experts routed eight at a time in this configuration. The context window is 1,000,000 tokens. The modality is text only, and the card says so without hedging. It arrives in three precisions, BF16, FP8 and NVFP4, plus a 2B DFlash2 speculative-decoding drafter. Its chat template runs a modified Harmony format with analysis and final channels and a reasoning-effort setting with six named levels. Its four repositories were published on the same day, 16 October 2026, and Artificial Analysis logged the release on its model page as 16 October 2026.

Read those two paragraphs again and the shape of the thing appears. Liquid AI published a component. Reflection published a finished system with a wire format and an effort dial. Both are called "open," and the word is doing two different jobs.

The licence is where the two are least alike

This is the comparison you cannot get from a parameter count, and it is the one with the most practical consequence.

Beam 1.0 is Apache 2.0. That is a standard permissive licence with no user caps, no monthly-active-user threshold and no separate commercial tier. Download it, fine-tune it, redistribute the derivative, sell the result. The card also carries a Responsible Use Guide and a separate safety technical report, which are policy documents rather than licence restrictions — they govern acceptable use, not the weights.

LFM2.5 2.6B Base is not. The Hub tags it license:other and names the licence as LFM Open License v1.0. The practical difference is not that one is permissive and the other is not — Liquid AI's licence is written for open use — but that the two licences have to be read differently, and only one of them can be summarised in a sentence. If your plan involves redistributing a derivative commercially, the Apache 2.0 side of this page requires no analysis. The LFM side requires that you read the licence, because a custom licence is by definition a document someone wrote on purpose.

That asymmetry runs the other way on the other axis. Liquid AI's checkpoint is 2.698 billion parameters of dense text model, which means the licence question is attached to something you can put on a single consumer GPU or an edge device. Reflection's Apache 2.0 grant is attached to a 501-billion-parameter model whose bfloat16 weights run to roughly a terabyte and whose four-bit variant still lists at 255B. A permissive licence on a model you cannot host is a different kind of freedom from a custom licence on a model you can.

What the base model can and cannot do

An article about a base checkpoint has to say this plainly, because it is the most common way people waste an afternoon. LFM2.5 2.6B Base has not been instruction-tuned. It will continue your text. It will not answer your question, follow your format, call your tool or stop when you want it to. It is a starting point whose behaviour is whatever you train into it, and Liquid AI says so in the model description rather than leaving it to be discovered.

That is not a defect and it is not a smaller version of Beam 1.0. The 34-trillion-token pretraining run and the 30-layer hybrid architecture are the product; the missing alignment is the point, because alignment is task-specific and the lab that ships a base model is deliberately handing that decision to whoever owns the task. Sixteen derivatives on the Hub is what the arrangement looks like in practice.

The consequence for evaluation is that there is no benchmark table to compare. Beam 1.0's card carries a full one — SWE-bench Verified 81.6% on BF16, Terminal Bench v2.1 80.1, GPQA Diamond 90.6, AIME 2026 97.7, LiveCodeBench 88.3, MCP Atlas 80.8, IFBench 79.6, all vendor-reported and none reproduced by a third party — because it is a post-trained model built for named workloads and that is the evidence its buyer asks for. LFM2.5 2.6B Base has none, by design. There is nothing to score: the checkpoint's quality is what your fine-tune makes of it, and the honest comparison point is a training run, not a leaderboard row.

Where the money goes

Neither of these is on OrcaRouter's catalogue. What we do route is the served world around them: more than 200 models behind one OpenAI-compatible key, provider list price passed through at 0% markup, with automatic failover across providers so a route that degrades reroutes instead of failing. That is worth naming here for a specific reason, because the two models on this page sit at opposite ends of the same procurement question.

LFM2.5 2.6B Base costs nothing to download and everything to operate, in the sense that the cost is a training run and a device. Beam 1.0 costs nothing to download and everything to operate, in the sense that the cost is serving a 501B MoE and, at time of writing, Reflection publishes no API rate. Artificial Analysis records the price fields on Beam's page as unavailable and notes there is no publicly available API. Twelve days after the announcement, the model with the 501-billion-parameter benchmark table still has no number on it that a finance team can use.

For a reader whose actual requirement is inference this week on a workload neither model is shaped for — a summariser, a classifier, a chat surface — the decision is not between these two. It is between fine-tuning 2.698 billion parameters yourself and renting something that is already aligned and priced, and the second option is the one you can measure before you commit.

Which one you are actually looking for

Take LFM2.5 2.6B Base if your problem is narrower than a general assistant. The reason to pick a small base checkpoint is that you have a task, you have data, and you would rather own the artefact than rent a call. The 131,072-token context is generous for the size, the 16-language coverage is real, the hybrid convolution stack is chosen for on-device throughput, and the licence lets you distribute what you build. What you are signing up for is the fine-tuning pipeline, the evaluation harness and the alignment decisions — none of which the checkpoint supplies.

Take Beam 1.0 if your task is one of the ones its card names: coding, tool use, repository-scale reasoning, long-horizon agentic execution. The evaluation table is aimed straight at that work, the 1M-token context matches it, the FP8 and NVFP4 builds make the hardware conversation more tractable than 501B suggests, and the Apache 2.0 licence clears a commercial derivative. What you are signing up for is serving infrastructure, an always-on reasoning budget the template will not let you switch off, and a cost estimate you have to build yourself because the vendor has not published one.

The two models are not in competition, and the comparison is worth making precisely because it shows how little "open weights" tells you on its own. One publisher handed over a component and a training budget; the other handed over a system with a wire protocol and a licence you can paste into a legal review. Reading which one you were offered before you download it is most of the work.

What would change the picture

A published Beam 1.0 API price, or a hosted FP8 deployment with a rate card, would move that model from the "build an estimate" column into the "compare the numbers" column, and it is the single missing row on its documentation. An independent benchmark on Beam's weights — now possible, since they are public — would give the vendor table a referee, which is what LFM2.5 2.6B Base can never have because there is nothing yet to score. Until one of those lands, the accurate summary is that a 2.7-billion-parameter base checkpoint and a 501-billion-parameter post-trained MoE are both free to download, both expensive to operate in completely different currencies, and neither is a substitute for the other.

Comparisons like this one are also where the failure modes show up: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.