Two things happen together. LiquidAI publishes two encoders, LFM2.5-Encoder-230M and 350M, that process long contexts on CPU at a speed 3.7 times faster than ModernBERT-base. Anthropic, for its part, formalizes Claude Fable 5 as a permanent model in Max and Team Premium plans, at 50% of the limits, closing weeks of temporary extensions (as we reported on July 18).

Why it matters

If you have tasks running all day (intent routing, policy filters, PII detectors), a fine-tuned encoder on CPU costs a fraction of a cloud LLM and completes in seconds. LFM2.5-Encoder-230M scans an entire contract or long support thread in under 30 seconds on a laptop. Where ModernBERT-base takes a minute and a half.

On the subscription front, Fable 5’s stabilization gives Claude Max or Team Premium users a reliable default model. No more monthly renegotiations with OpenAI over GPT-5.6: the choice shifts from marketing to cost per verified output.

If you want to try it

The models are on Hugging Face with live demos in CPU-only spaces: zero-shot prompt routing, policy linting, PII detection in 16 languages.

In detail

The inference market is bifurcating. On one side, frontier models (GPT-5.6, Claude Fable 5) compete on power and subscription cost. On the other, specialized and small models that run on hardware you already own solve repetitive tasks without the cloud.

LFM2.5-Encoders comes from LiquidAI, not Anthropic. They’re two encoders (230M and 350M parameters) built on the LFM2 architecture, the same as the retrievers released last month. The shift from decoder to encoder happens through three changes: bidirectional attention (each token looks both left and right), non-causal convolutions with symmetric padding, and masked language modeling with 30% of tokens hidden during training.

Training has two phases. First, general competence on short context (1,024 tokens), then adaptation to long context (8,192 tokens) on multilingual, legal, and factual data. On GLUE and SuperGLUE benchmarks, LFM2.5-Encoder-350M ranks fourth out of 14 models, beating encoders ten times larger. The 230M beats ModernBERT-base and every EuroBERT, being smaller than most of them.

CPU speed is the decisive point. At 8,192 tokens, ModernBERT-base takes over 90 seconds per forward pass. LFM2.5-Encoder-230M takes 28: 3.7 times faster. On GPU the advantage exists but is smaller, and only shows up above 2,000 tokens.

Live demos on Hugging Face show what you can build: zero-shot prompt routing (define lanes as free text, the model assigns each prompt to the right one in a single pass), policy linting (check text against company rules written in natural language), token-by-token spell checking, and PII detection with 40 categories of personal information in 16 languages.

For those building agents, this means text comprehension tasks (routing, classification, extraction, scoring) can run locally, without network latency and without per-token cost. The natural placement is in the pre-processing layer: first the encoder decides if the large model is needed, then calls it only when necessary (like Cactus Hybrid does with confidence probes).

Fable 5’s stability in Anthropic plans responds to a different pressure. GPT-5.6 imposed an update pace that forces subscription managers to renegotiate every month. Fixing Fable 5 as permanent, even at 50% of the limits, gives enterprise customers a stable reference point. The trade-off is clear: less capacity than the top tier, but predictability.

The limits: the benchmarks are declared by LiquidAI itself, with open-source framework but internal measurements. The models are new, nobody has stressed them in production on real workloads for weeks yet. CPU speed depends on the specific processor: the numbers cited are averages, not guarantees for every hardware. The claim about Fable 5 permanent at 50% of limits is confirmed by the July 18 piece but has no primary source among the items collected this round.