LPCompact-Qwen3.8-27B

LPCompact 0.1 is a learned state compaction model for Qwen3.8-27B, available through LatentPort's state.compact() operation.

It selects useful attention-memory entries directly from a running model's state. Recurrent state, convolution state and the logical continuation clock are preserved. The operation uses no text summary or historical token replay.

Model and runtime

This repository contains the trained compaction selector for Qwen3.8-27B: one model.safetensors file, its configuration, model card and license. Qwen base-model weights, including GGUF files, are obtained separately.

The shared Python runtime is maintained in LatentPort on GitHub. Each trained state operation has its own Hugging Face repository, named by operation and model compatibility. This model supplies compaction for Qwen3.8-27B through that common runtime.

One checkpoint across BF16 and GGUF

The same 393,473-parameter selector is used with Qwen3.8-27B BF16 and GGUF weight quantizations. Base-model weight precision does not change the selector's architecture or weights. LatentPort adapters supply the common BF16 attention-state layout.

Runtime qualification passed at 64K and 128K with BF16, Q4_K_M and Q8_0, including continued generation, save/load, forks and repeated compaction. See the runtime validation record.

Use with LatentPort

Install the LatentPort runtime release and follow the BF16 or GGUF setup instructions.

import latentport as lp

with lp.load_model(backend="bf16") as model:
    state = model.prefill(history)
    state = state.compact()
    answer = model.session(state).ask("Continue the task.")

For GGUF, use backend="gguf" and supply model_path with the chosen Qwen3.8-27B GGUF. Both formats automatically load the same selector.

The returned value is an ordinary immutable LatentPort State, supporting fork, save/load, append and continued generation. The original snapshot remains valid.

Compaction policy

The default policy retains one eighth of attention KV positions, with a 1,024-position floor. A 4x setting is also available. Normal activation starts at 65,536 physical KV positions. At 128K, 8x selection reduces BF16 attention tensor storage from 8 GiB to 1 GiB. Base weights and fixed recurrent state retain their existing representation.

The selector reads three attention value layers and assigns a utility score to each retained position. Its span policy preserves the first and last 512 positions and selects the remaining memory within the requested budget.

Artifact and version

LPCompact 0.1 uses FP32 selector weights, a 3,072-feature input, a 128-unit hidden layer with GELU, and one output score. config.json records its architecture, policy and hashes. The v0.1 tag identifies the original release checkpoint. The model weights are unchanged by the repository rename.

Canonical repository: latentport/LPCompact-Qwen3.8-27B.