
Clef-Flash vs LFM2.5 2.6B Base: Two Models Called "Small," Three Orders of Magnitude Apart
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 186 tok/s
- openaiNEWOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 86 tok/s
- openaiNEWOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 151 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 192 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 86 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 705 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 246 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 97 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 138 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 85 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 102 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 70 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 38 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 · 231 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 101 tok/s
The clef-flash build most people actually download is 11 GB. LiquidAI/LFM2.5-2.6B-Base runs in under 2.5 GB of memory. Both models get described as small, fast and efficient by the labs that made them, and both descriptions are accurate — which is exactly why the comparison is useful, because the two words are doing completely different work. Cloudflare/clef-flash is a 9B multimodal decision model that answers typed questions with calibrated probabilities and never writes a token; it needs an H200-class GPU to hit the 38.8 ms median latency on its own card, and Cloudflare's pitch for it is edge inference inside their network. LFM2.5 2.6B Base is a pre-trained text checkpoint Liquid AI published on 1 August 2026 so that someone else can fine-tune it, and its intended deployment target is a laptop.
Neither of them is a language model you would put in a chat window. Both are cheap. Choosing between them is not really a choice about quality — it is a choice about what your hardware budget is and whether the thing you are building already knows what its answers look like.
The footprint conversation, in numbers
Start with what the machines actually need, because this is the axis on which the two are furthest apart.
• Parameters — Clef-Flash is a frozen Qwen3.5-9B backbone with its vision encoder plus a 243.5 MB joint schema head; LFM2.5 2.6B Base is 2.69 billion parameters, text only.
• Disk and memory — the community clef-flash build is 11 GB; the full Hugging Face repository is about 18.8 GB across four backbone shards plus the head. LFM2.5 2.6B Base is quoted at under 2.5 GB of memory for inference, and quantised GGUF, ONNX and MLX builds are published alongside it.
• Latency — Clef-Flash: 38.8 ms median, 122.4 ms p95, measured by Cloudflare on a single H200. LFM2.5 2.6B, the post-trained sibling of this base checkpoint, runs at 220 tokens per second on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU.
• Context — 65,536 tokens for Clef-Flash per Cloudflare's Workers AI documentation; 131,072 tokens for LFM2.5 2.6B Base.
• Licence — Apache-2.0 for clef-flash; the LFM 1.0 licence for LFM2.5, which is open weights but not an OSI-approved open-source licence, and the difference matters if you are shipping a product rather than experimenting.
Those two latency lines are not measuring the same unit and you should not subtract them. But read together they say the obvious thing plainly: Clef-Flash is a datacentre component whose vendored runtime also happens to sit on an edge network, and LFM2.5 2.6B is a component for the machine on your desk.
Why the base checkpoint is a different kind of object
The word "Base" in the name is load-bearing, and Liquid AI's own card is unusually candid about it. LFM2.5-2.6B-Base is the pre-trained text-only checkpoint, 30 layers arranged as 22 double-gated short convolution blocks and 8 grouped-query attention blocks, trained on 34 trillion tokens over a 128,000-token vocabulary in sixteen languages. It is "only recommended for tasks that require heavy fine-tuning" — language-specific work, domain-specific assistants, training on proprietary data. Its card says so in as many words.
Clef-Flash is the opposite of that. It is finished. The training is done, the schema head is bolted on, the API contract is published, and the model is hosted on the vendor's own edge network at a listed price. You do not tune it; you call it, or you download it and call it locally. There is a fine-tuning product attached to the family — Cloudflare launched one alongside the models — but using clef-flash does not require it, and that is the point.
So the honest framing is that these are two stages of two different pipelines. One is an ingredient and one is a component. If your task is "I have 40,000 labelled support tickets and I want a fast domain-specific classifier that runs on our own boxes", LFM2.5 2.6B Base is a plausible starting point and clef-flash is not, because you cannot fine-tune clef-flash's decision head into a language capability it does not have — it has no language output path at all. If your task is "I need to route a request into one of six buckets in under 50 milliseconds with a probability I can threshold", clef-flash answers that question out of the box and LFM2.5 2.6B Base does not answer it in any form.

The part that is genuinely shared
There is one real overlap, and it is worth being precise about it because it is where the comparison stops being trivial.
A post-trained generalist — LFM2.5 2.6B, or any small instruct model — can be prompted to return a structured classification. Give it a JSON schema, ask for a label and a confidence, parse the output. People do this constantly. The question is what you get when you do.
What you get is a token sequence that happens to contain a label you can extract, plus a number the model wrote about itself. That confidence is a sampled token, not a softmax over candidate answers, and it does not mean what a calibrated probability means. Clef-Flash's card is explicitly careful about the analogous figure in its own API — confidence "shows how concentrated the probabilities are. It isn't the chance that the answer is right" — and it can afford to be, because the underlying quantity is a real distribution over the options you supplied. A prompted generalist has no such distribution to expose. It has a sentence that says 0.85.
Whether that distinction is worth 11 GB and an H200 depends entirely on what you do with the number. If the label goes into a dashboard, a prompted 2.6B model is fine and cheaper by every measure. If the label gates an irreversible action — issuing a refund, blocking an account, escalating a security incident — the difference between a calibrated score and a plausible-looking string is the difference between a threshold you can justify to an auditor and a threshold you cannot.

Where you can get each of them
Neither model is on OrcaRouter. Our catalogue does not carry clef-flash, and it does not carry LFM2.5 2.6B Base or any other Liquid AI checkpoint — you can verify both by hitting our model API and getting a 404. Nothing in this article is an availability claim from us.
Clef-Flash is reachable on Cloudflare's Workers AI at $0.09 per million input tokens, or as Apache-2.0 weights you run yourself. LFM2.5 2.6B Base is on Hugging Face with quantised builds for llama.cpp, ONNX Runtime and MLX, and Liquid AI's own playground and documentation for the family. If you want a decision model that is live on our catalogue today, the one that is there is TypeSafe's Jev 1.13 — same request shape, 65,536-token context, $0.042 per million input tokens — and clef-flash was built to be a drop-in for it, which means trying either one is a config change rather than a rewrite. Both call the same endpoint class, so if you are already routing through one key with provider list price passed through at no markup, the comparison is an experiment you can run this afternoon rather than a migration you have to plan.

Which one belongs in your stack
The decision tree is short. If your answers are a finite list you can write down in advance, and you need a probability that means something, and you have the hardware or the willingness to pay per call — clef-flash is the finished component and LFM2.5 2.6B Base is not a candidate. If your answers are open-ended, or the thing you actually need is a small language model you can shape with your own data and run on a laptop, LFM2.5 2.6B Base is the starting point and clef-flash cannot help you at all.
The case where people get this wrong is the middle: building a bounded classifier by prompting a small generalist, because it is the path of least resistance and it works in the demo. It works in the demo. It stops working the moment you need the number to be a number, and by then the schema, the prompts and the parsing code are all load-bearing.
One caveat on the clef-flash side, stated plainly because it is the vendor's own weakest figure: on CLINC150 with out-of-scope handling, where the right answer is sometimes "this fits none of your categories", clef-flash scores 66.8 macro-F1 against 97.4 for its larger 27B sibling. A bounded schema is a strength until the input does not fit it, and the model will classify anyway. Whichever route you take, build the residual class deliberately and audit what lands in it.
