Sakana Namazu vs LFM2.5-2.6B-Base: The Real Price of Free Weights

LFM2.5-2.6B-Base costs $0.00 to download. Sakana Namazu costs $0.95 per million input tokens and $4.00 per million output tokens. Read those two price tags and you would conclude the comparison is over before it starts — but the actual price difference runs the other way. LFM2.5-2.6B-Base, Liquid AI's 2.69-billion-parameter pre-trained checkpoint released on Hugging Face on August 4, 2026, is not a product: no inference provider hosts it, it has zero published evaluations, and its own card calls it material for fine-tuning. Sakana Namazu, by contrast, is a finished, served, tooled API. The question is what you are actually willing to pay — in money, or in work.

What each one actually is

Sakana Namazu is a post-trained Kimi K2.6 tuned for Japanese business language, wrapped in an OpenAI-compatible API with web search and code execution built in. You change a base_url, set a key, and it works. LFM2.5-2.6B-Base is the raw pre-training checkpoint behind Liquid's LFM2.5-2.6B family — the stuff the post-trained, agentic sibling is built from. Liquid's own card is explicit that the base exists for heavy fine-tuning, and its benchmark table belongs to the post-trained version, not this one.

The spec sheet

• Price — Sakana Namazu $0.95 / $4.00 per 1M tokens vs LFM2.5-2.6B-Base $0 to download, self-serve everything

• Size — Namazu inherits Kimi K2.6's 1T / 32B-active MoE vs LFM2.5-2.6B-Base 2.69B dense hybrid (22 conv blocks + 8 attention layers)

• Context — Namazu not disclosed vs LFM2.5-2.6B-Base 128K tokens (131,072 per config)

• Hosting — Namazu served by Sakana vs LFM base hosted by nobody — you build the serving stack

• Evaluations — Namazu vendor-reported vs LFM base zero published scores

• Language — Namazu Japanese-business tuned vs LFM base multilingual (16 languages, 128K-token vocabulary)

Doing the total-cost math

"Free weights" is a real price, it is just denominated in labor. To get LFM2.5-2.6B-Base to do a job you need: a fine-tuning run (Liquid's own math — roughly $10 of tokens for a 100K-example supervised set, about $100 for 1M examples), an inference stack (vLLM, SGLang, or llama.cpp on a device you own), and ongoing serving. The payoff is that the meter stops: after the build, every token is free, which makes the base the runaway winner for high-volume, always-on, latency-tolerant workloads. Namazu's meter never stops, but it starts instantly — no training run, no GPU, no ops. For a one-off integration, or a workload where your team's time is the expensive line item, the "free" model is the expensive one.

Licenses that bind

The fine print cuts against the free model too. LFM2.5-2.6B-Base ships under the LFM Open License v1.0, which caps commercial use at $10 million in annual revenue — and applies that cap to derivative works, meaning your fine-tune inherits it. Cross the threshold and you owe Liquid a commercial license. Sakana Namazu has no such cap, but it is closed, and its data-handling defaults (training on your inputs unless you opt out, no guaranteed Japan-only processing) are the price of the hosted convenience. Neither is free in the sense the sticker implies.

The agent gap

The sharpest difference is what the two models are for. LFM2.5-2.6B-Base is a base — the raw substrate from which Liquid's post-trained agentic model is derived. It is not an agent; it is not even an instruct model. Sakana Namazu is an agentic product: built-in web search, code execution, and a demonstrated self-directed loop (Sakana's demo has it planning a market-research report, searching, cross-checking, and writing it unattended). If the job is "run a 2.6B model you tuned for one narrow, high-volume task," the base wins. If the job is "have an agent do knowledge work," Namazu is the only one of the two that shows up with tools.

So which is cheaper?

The honest answer is that the two are not substitutes — the base is a component, and Namazu is a product — so the real decision is what your stack needs next. In practice neither lives alone: a fine-tuned 2.6B specialist ends up sitting next to a generalist agent model, and a routing layer is how you keep that multi-model reality sane. OrcaRouter's model fusion and routing DSL let you compose several models into one call — the pragmatic middle path between "serve a 2.6B checkpoint yourself" and "buy one vendor's API" — and, because we pass provider prices through at zero markup, the billing stays predictable while the mix changes. Start with Sakana Namazu if you want a working Japanese-business agent today; start with LFM2.5-2.6B-Base if you are prepared to build the product yourself and amortize the build across billions of free tokens.