
Aion-1.0-Instruct-2.3B vs LFM2.5 2.6B Base: Almost the Same Size, Completely Different Deliverables
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 184 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 228 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 131 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 77 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 69 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 208 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 311 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
On paper this is the tightest matchup in the small-model field right now. Aion-1.0-Instruct-2.3B is 2.29 billion parameters; LFM2.5 2.6B Base is 2.69 billion. Both are dense, both are text-only, both are open weights, both target on-device deployment, and both arrived without a launch event — Microsoft's on 18 November 2026, Liquid AI's base checkpoint on 4 August 2026. If you lined them up by specification you would conclude they are competitors.
They are not. One of them is finished. The other is a starting point that has deliberately not been finished, and Liquid AI says so on the card: the Base checkpoint "is only recommended for tasks that require heavy fine-tuning." That single sentence reorganises everything else in this comparison, because a pretrained base and an instruction-tuned assistant are not two products in the same category — they are two stages of the same pipeline, sold separately.
The Aion-1.0-Instruct-2.3B figures below come from Microsoft's model card and repository configuration, read directly. The LFM2.5 2.6B Base figures come from Liquid AI's model card and its accompanying blog post. Neither model has an independent evaluation; where a number is vendor-reported it is labelled, and where a number does not exist at all, that is said rather than filled in.
What "Base" removes from the comparison
LFM2.5 2.6B Base is the pretrained checkpoint from which every LFM2.5 2.6B variant was created. It was trained on roughly 34 trillion tokens across 16 languages, and it will continue text rather than follow an instruction, because instruction tuning lives in the sibling checkpoint. Load it and ask it a question and it will do what a language model does without post-training: extend your prompt in a plausible direction.
That makes about half of the usual comparison invalid. There is no point asking which model follows instructions better, because one of them has not been taught to. There is no point comparing refusal behaviour or chat formatting. What remains is a narrower and more useful question: if you are building an assistant rather than buying one, is Microsoft's finished 2.29B a better starting point than Liquid AI's raw 2.69B — or is it a worse one, because its tuning has already committed it to a shape you may not want?
• Parameters — 2.29B with tied embeddings for Aion; 2.69B for LFM2.5 2.6B Base.
• Architecture — Aion is a conventional transformer with grouped-query attention and a hybrid 21-sliding-plus-7-full attention pattern; LFM2.5 is a hybrid of 22 double-gated short-convolution blocks and 8 GQA layers across 30 layers total.
• Training volume — Microsoft does not publish a token count for Aion; Liquid AI publishes roughly 34 trillion for LFM2.5.
• Context — 32,768 tokens for Aion, 131,072 for LFM2.5 2.6B Base, with a dedicated 128K long-context extension phase in mid-training.
• Vocabulary — 200,029 for Aion; 128,000 for LFM2.5, doubled in place from the previous generation's tokenizer rather than retrained.
• Licence — MIT for Aion; the LFM Open License v1.0 for LFM2.5 2.6B Base, which is not an OSI licence and needs reading before you ship.
• Post-training — an instruction-tuned model with a chat template and tool-calling format, against a pretrained checkpoint with neither.
The architecture choice that explains the context gap
The fourfold difference in context window is not an accident of training budget. LFM2.5's design replaces most attention layers with short-convolution blocks, which is a bet that sequence mixing does not need full attention at every layer to work — and that bet is what makes a 131K window affordable on a phone. Liquid AI's blog is explicit that mid-training included a dedicated 128K context-extension phase, so the long window was trained for, not configured on afterwards.
Aion goes the other way. Twenty-one of its 28 layers use a 2,048-token sliding window and only seven use full attention, with YaRN scaling of factor 16 applied to those seven. The published context is 32,768, and the configuration's original_max_position_embeddings of 2,048 is the tell: the window is reached by positional interpolation from a short native range. This is a normal technique. It is also unmeasured here — Microsoft published no long-context evaluation, and the model is a day old, so no one else has run one either.
Practical translation: LFM2.5 2.6B Base comes with a long context you can verify against a documented training phase. Aion comes with a shorter one you should test rather than assume.
What Aion gives you that a base checkpoint cannot
The finished-model advantages are concrete and boring, which is what makes them worth listing. Aion-1.0-Instruct-2.3B ships a chat template with a default system preamble, an optional user-supplied system message appended to it, and template variables for model identity, knowledge cutoff and current date. It ships a documented tool-calling format. It has been safety-aligned to Microsoft's standard, and the card reports XPIA, DIC and ECI as the safety metrics it intends to improve — naming the gaps rather than claiming they are closed.
It also ships eleven benchmark scores on a named harness, which is more than LFM2.5 2.6B Base offers. The base card prints architecture and training facts and no evaluation at all, which is the correct choice for a checkpoint that is not meant to be scored on tasks it was not tuned for. If you need a number to put in a design document, Aion has one: IFEval 83.90, MMLU 55.15, GSM8K 74.91, MMLU-Pro 33.35, GPQA 28.18, among others, all vendor-run on lm-evaluation-harness or the benchmark's own implementation.
But read the caveats on Aion's own card before you treat that as a capability ceiling. It says the model "was not trained to excel at world knowledge", that a model of this size "can miss fine details, provide misstatements, or show inability to follow complex multi-step reasoning", and that multilingual data was about 6% of training across eight languages with quality varying by language. That is an unusually frank limitations section, and it describes a model tuned for summarisation, rewriting, chat and operating-system tasks rather than for open-ended reasoning.
What the base checkpoint gives you that a tuned model cannot
Control, and a clean starting loss surface. If your product needs a model that writes in a house style, answers in a domain vocabulary, or emits a strict structured format, post-training on a base checkpoint gets you there without fighting someone else's alignment pass. Liquid AI's card is direct about the intended uses: heavy fine-tuning, language-specific assistants, domain-specific assistants, training on proprietary data, and experimenting with novel post-training approaches.
It also gives you a licensing decision to make rather than inherit. LFM2.5 2.6B Base is released under the LFM Open License v1.0 — a custom licence, not Apache 2.0 and not MIT. Its card lists the licence as "other" with the name "lfm1.0". Whether that is a problem depends entirely on your distribution model, and it is the kind of thing that surfaces late in a legal review if nobody reads it early. MIT, by contrast, is about as short as licences get.
And there is a tooling argument. LFM2.5 shipped with day-one support across Transformers, vLLM, SGLang, llama.cpp, MLX and LM Studio, plus published fine-tuning notebooks for Unsloth and TRL covering continued pre-training, SFT, DPO and GRPO. Aion Instruct needs trust_remote_code=True and transformers 5.x, has no GGUF file in its repository despite the card listing GGUF as a format, and its quantised builds are promised rather than published. For a model you intend to fine-tune, the ecosystem around it is not a detail.
Two ways to end up in the same place
Consider the two realistic paths for a team that wants a small on-device assistant and has a corpus of their own data.
Path one: take Aion-1.0-Instruct-2.3B, evaluate it on your task as shipped, and fine-tune only if it falls short. You start with a working assistant, a chat template and a tool-calling format, and your first experiment is cheap because the model is already doing the job badly rather than not at all. The risk is that you are correcting someone else's alignment rather than teaching from scratch, and at 2.29B there is not much capacity to correct with.
Path two: take LFM2.5 2.6B Base, run continued pre-training on your domain, then SFT. You get a model shaped by your data rather than adjusted toward it, with a 131K window available from the start and a well-documented training toolchain. The cost is that you own the whole post-training effort, including the safety behaviour Microsoft already handled.
The third path is the one most teams actually take: start on path one to get a baseline, and switch to path two only if the baseline plateaus. That is a rational order, and it is also an argument for not signing a long contract on either model before you have run your own data through both.
Once you have two checkpoints in a product, the operational surface starts to sprawl — separate serving configs, separate failure modes, separate upgrade cadences. OrcaRouter's contribution at that layer is a single endpoint over 200+ models at 0% markup with provider list price passed through, automatic failover across providers, and a routing DSL that can express a fallback chain instead of a hard-coded target. Neither Aion-1.0-Instruct-2.3B nor LFM2.5 2.6B Base is in that catalogue — both are self-hosted weights, and Aion has no hosted API anywhere — so what the router buys you is the tier above them: somewhere to send the requests your on-device model should not be handling.
The decision, compressed
If you want something running this week that answers questions, take Aion-1.0-Instruct-2.3B. It is finished, it is MIT, it is 4.6 GB in bfloat16, and its limitations are documented in its own words rather than discovered by you.
If you want to build a model rather than adopt one, take LFM2.5 2.6B Base. It is 5.4 GB in bfloat16, it carries a 131K window that was trained rather than scaled, it has a real fine-tuning toolchain, and it has no opinions about what your assistant should sound like because it has not been given any.
What you should not do is treat them as interchangeable because the parameter counts are close. A 2.69B base and a 2.29B instruct model differ by less than the rounding error on most spec sheets and by an entire post-training pipeline in practice, and choosing between them is a decision about who does the work.

Liquid AI's card lists the numbers that matter for a fine-tuning target — 2.69B parameters, 30 layers, 34 trillion training tokens, a 131,072-token window — and no evaluation table, because a base checkpoint is not meant to be scored on tasks it was not tuned for.
The open questions
For Aion: whether the quantised builds Microsoft promised arrive, whether the long-context window holds up under a real needle test, and what happens to this checkpoint when the further-optimised Windows version lands by the end of January 2027. The card presents this upload as a preview of that model, which means its successor is already scheduled.
For LFM2.5 2.6B Base: whether the community builds enough on it to make its fine-tuning recipes the reference path, and how the LFM Open License is read by legal teams outside the usual permissive-licence comfort zone. Both are answerable with time, and neither is answerable by a benchmark table.

The Aion card states the intended use in one line: the foundational SLM for on-device Windows tasks, with the OS integration pencilled in for January 2027.

Nearly identical parameter counts, different deliverables: instruction-tuned and MIT against pretrained and custom-licensed, with the context window running the other way.
