
Aion-1.0-Instruct-2.3B vs Aion 3.5 Mini: Two Models Named Aion, One Company Apart
- 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 · 223 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 · 78 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
Microsoft's Aion-1.0-Instruct-2.3B appeared on Hugging Face on 18 November 2026 with no launch post, no blog, and three likes. Aion 3.5 Mini has been on sale since 23 September 2026 at $0.70 per million input tokens, sold by a company called AionLabs. They share a name and nothing else — not a licence, not an architecture, not a business model, and not a single line of code. Anyone searching for "Aion" right now will hit both, and picking the wrong one costs either a wasted download or a wasted invoice.
What follows separates the two carefully. Everything about Aion-1.0-Instruct-2.3B comes from the model card and configuration file Microsoft published with the weights, read directly; it is one day old at the time of writing and no independent evaluator has touched it. Everything about Aion 3.5 Mini comes from AionLabs' published model list and the served catalogue on its own API. Where a vendor reports a number, that is stated; where nobody has measured something, that is stated too.
The name is the only thing they share
AionLabs is a model vendor that sells hosted inference. Microsoft's Aion comes out of the Windows Applied Sciences Group and is described on its card as "the foundational Small Language Model (SLM) in the Windows ecosystem, specifically optimized for on-device AI". One is a product you rent by the token. The other is a weights release that Microsoft frames, in its own words, as "an introduction to what's coming and a way to maximize transparency with our users".
The AionLabs side does not even publish a parameter count. Its catalogue entry describes Aion 3.5 Mini as "a multi-model system built on the GLM family of models" in which "multiple specialized models contribute to each response" — an ensemble behind an endpoint, not a single checkpoint you could download. There is no weight file to compare against Microsoft's 2.29 billion of them.
• Parameters — Aion-1.0-Instruct-2.3B is ~2.29B with tied input/output embeddings; Aion 3.5 Mini publishes none, because it is a routing layer over several models rather than one network.
• Weights — MIT-licensed bfloat16 safetensors you can download today versus a closed hosted service.
• Context — 32,768 tokens versus 262,144.
• Maximum output — the AionLabs catalogue caps a response at 32,768 tokens; Microsoft publishes no separate output ceiling, so the ceiling is the window.
• Modality — text in, text out on both sides. Neither is multimodal.
• Price — zero for the Microsoft checkpoint, $0.70 / $0.18 cached / $1.40 per million tokens for Aion 3.5 Mini.
• Reasoning controls — Aion 3.5 Mini exposes low, high and max effort levels with high as the default; Aion-1.0-Instruct-2.3B has no thinking mode at all and answers in one pass.
What Microsoft actually shipped, and what it did not
The card's architecture table is unusually specific for a quiet upload: 28 layers, hidden size 2,048, an 8,192-wide SiLU MLP, grouped-query attention with 24 query and 8 key/value heads at head dimension 128, and a hybrid attention pattern of 21 sliding-window layers with a 2,048-token window plus 7 full-attention layers. The vocabulary is 200,029 tokens and the checkpoint is bfloat16.
The context claim deserves a second look. The published window is 32,768 tokens, but the configuration's RoPE block records original_max_position_embeddings of 2,048 with a YaRN scaling factor of 16 applied to the full-attention layers only. Sixteen times 2,048 is 32,768, which suggests the long window is reached by positional interpolation over a much shorter native range rather than by training at length. That is a legitimate engineering choice and a common one at this size. It is also the kind of detail that decides whether a 30K-token document summarisation actually holds up, and nobody has published a long-context evaluation of this model.
Two other gaps are worth naming plainly. The card lists GGUF among its library names, but the repository's file listing contains safetensors, tokenizer files, custom modelling code and a chat template — no GGUF. The card says the unquantized model is being published first "to be followed by quantized versions", so the format is promised rather than present. And the Windows story is explicitly future tense: "a further optimized version of Aion Instruct, coupled with Microsoft's content moderation model, will become the Windows OS SLM by the end of January 2027". The upload is a preview of that model, not the model Windows will ship.
What is present is a working instruction-tuned artefact-in-a-repo: a chat template with system, user and assistant turns each closed by a dedicated end-of-turn token that doubles as the EOS token, and a documented tool-calling path that injects function schemas into the system prompt and parses calls back out of wrapped tags. It also requires trust_remote_code=True, because the modelling code ships with the weights rather than living in transformers.
The number that separates the two products
Aion 3.5 Mini has no published benchmark. None. AionLabs' model list gives price, context, output ceiling, modality and reasoning levels, and stops there. There is no evaluation card, no third-party index entry we could find, and no head-to-head against anything.
Microsoft went the other way. Aion 3.5 Mini's absence of a scoreboard is not a gap you can fill from elsewhere; the Aion-1.0-Instruct-2.3B card publishes eleven benchmark results with the harness and judge model named, and every one of them is dated, harnessed and attributable. MMLU 55.15, IFEval 83.90 and MMLU-Pro 33.35 are the three worth carrying in your head, because they bracket what the model is: competent at following a specified format, mediocre at recall and multi-step reasoning. Ten more rows sit on the card for anyone who wants the full picture.
Two footnotes on that table matter more than the table. GSM8K carries the note "loose matching", which flatters the score relative to strict extraction, and IFBench is reported "prompt-level loose". A vendor that volunteers its own lenient scoring settings is telling you the numbers are not marketing, and the resulting table — a 33.35 on MMLU-Pro in particular — looks weak next to what comparable open models print. That is a credibility signal in the opposite direction from a glossy launch: it reads like a real harness run rather than a curated one. It is still vendor-reported and unreproduced.
Cost, once you stop paying per token
The economics invert completely depending on which side you pick. Aion 3.5 Mini bills $0.70 per million input tokens and $1.40 per million output tokens, with cached input at $0.18. A million output tokens costs $1.40 and arrives with a 262K window and three reasoning levels. You pay nothing upfront and you own nothing.
Aion-1.0-Instruct-2.3B costs about 4.6 GB in bfloat16 — 2.29 billion parameters at two bytes each — and then costs whatever your hardware costs. At 8 bits that halves; at 4 bits it lands near 1.2 GB, close to the footprint of a mid-range phone model. Microsoft's target is exactly that: text and operating-system tasks on the device, with the data never leaving the machine. The trade is the opposite of the hosted model's: no per-token bill, no 262K window, no reasoning mode, and no vendor to call when it gets something wrong.
If you need both shapes at once, the plumbing question is how many contracts you are maintaining. On OrcaRouter one key reaches 200+ models with 0% markup — the provider's list price passed through, so a vendor price change is visible on our side the same day — plus automatic failover if a provider degrades and a routing DSL for composing models into one call. Neither of these two Aions is in that catalogue: Microsoft publishes weights, not an API, and AionLabs is its own vendor. So this is not a "call both from one place" pitch. It is the narrower point that a hosted tier you rent and a checkpoint you run on a laptop answer different questions, and mixing them only makes sense when each is doing the job it was built for.
Which one you actually want
Take Microsoft's Aion-1.0-Instruct-2.3B if the constraint is locality. A document that cannot leave the machine, an operating-system surface that has to work offline, a latency budget measured in tens of milliseconds, and a workload built from summarisation, rewriting, chat and light tool calls — that is the card's own description of the model, and the 32K window is sized for multi-page documents rather than for whole codebases. Go in knowing the model is one day old, that its long-context behaviour is unmeasured, that no quantised build exists yet, and that the 2.3B class will misstate facts and lose the thread on multi-step reasoning, as its own limitations section says.
Take Aion 3.5 Mini if the constraint is capability per request and you are willing to pay for it: a 262K window, optional reasoning depth, tool calling, and a vendor who will answer a support ticket. Go in knowing that you cannot see the weights, cannot count the parameters, and cannot check any claim about quality against a published number — because none exists.
The trap is the name. A search for "Aion 2.3B" or "Aion Instruct" will surface AionLabs pages; a search for "Aion 3.5" will eventually surface coverage of Microsoft's Windows SLM work. Write down which vendor you meant before you write the integration. The two are unrelated, and the identifier is the only thing that will tell them apart in a dependency file.

Microsoft's card is the only place the 2.29B figure, the 32K window and the January 2027 Windows target appear together. It is also where the benchmark table and its "loose matching" notes live.
What would change this comparison
Three things, in order of likelihood. A GGUF build would put Aion Instruct into llama.cpp and LM Studio and make the local story practical rather than theoretical — the card promises quantised versions and the repository does not have one yet. An independent harness run would turn Microsoft's table into a checkable claim; at three likes and zero downloads on day one, nothing like that exists. And the January 2027 Windows integration would change the audience entirely, from people who download weights to people who never think about the model at all.
On the other side, a published evaluation of Aion 3.5 Mini — any evaluation — would be news, because the current situation is a paid product with a rate card and no denominator.

AionLabs' catalogue is where the Aion 3.5 Mini rate card and the 262,144-token window come from. There is no benchmark column on it.

The scoreboard, with the asymmetry made explicit: one side ships numbers and no price, the other ships a price and no numbers.
