What the base model is
Qwen3.8-27B is a dense model from the Qwen Team: every one of its roughly 27 billion parameters is used for every token, unlike the mixture-of-experts models elsewhere in this catalogue. Its card describes it as a native vision-language model that reads images and video, with a thinking mode that is on by default and can be turned off per request.
The architecture is unusual. Of its 64 layers, 48 use Gated DeltaNet, a form of linear attention, and 16 use full attention, in a repeating pattern of three linear layers and one full layer. Linear-attention layers keep a fixed-size state instead of a cache that grows with every token, which is why this model’s KV cache is small for its size. The VRAM estimator counts only the 16 full-attention layers for that reason.
The card states a native context of 262,144 tokens, “extensible up to 1,000,000”. Hosted providers set their own limits, which in our data range from 32,768 to over a million tokens depending on the provider and the build.
The variants at a glance
huihui-ai’s abliterated weights are the most transparent option. The card says the release was made with abliteration, a weight edit that removes a “refusal direction” identified in the model’s activations, and that only layers 18 to 51 were edited in the current version. The multi-token-prediction head and the vision encoder were left untouched. An earlier revision kept only the first 15 layers unedited, and the card gives the command to download it. huihui-ai also publishes GGUF files and an Ollama build.
HauhauCS’s “Aggressive” GGUF release describes itself as a profile for “direct answers” with minimal preamble, ships the model’s native multi-token-prediction head plus its own speculative-decoding sidecar, and reports “0/465 refusals” on a test it does not publish. It does not say how the change was made. The creator itself suggests a “Balanced” release is the safer default for long-context agentic work when one exists.
Hugging Face lists many more Qwen3.8-27B variants, including Heretic builds and fine-tunes; several are served by Featherless under their own repo names. We catalogue the two above because their cards are readable and their creators have a record of earlier releases. That is not a judgment that the others are worse.
Which one for which need
- To run locally and inspect what changed: huihui-ai’s release, in GGUF form. You can compare its outputs with the base model on your own prompts, because the base weights are public under the same license.
- For a hosted API serving known weights: Featherless serves
huihui-ai/Huihui-Qwen3.8-27B-abliterated by name. Its published output rate for this model is much higher than its rate for the standard Qwen3.8-27B, and its listed context is 32,768 tokens.
- For a hosted API with long context on one bill: MolochAI’s
qwen3.8-27b-uncensored lists a 200,704-token context at its flat rate, but the provider does not say which weights it serves. Venice serves only the standard Qwen 3.8 27B in our data.
Prices for every option are in the table below and on the price table.
Limits and open questions
Neither variant publishes a measurement of capability after the change. Abliteration usually aims to leave knowledge intact, but the research behind it (Arditi et al., 2024) and later community work show that results vary by model and by how many layers are edited. If you rely on the vision or thinking features, test them: huihui-ai says the vision part was not modified, but the language layers that interpret images were.
Apache 2.0 allows commercial use and modification, with notice requirements when you redistribute. That covers the weights; it does not make every use lawful, and responsibility for outputs stays with you. See how the main licenses differ.
Ordered as catalogued. Method, layers and claims are as each creator states them; Atlas has not tested any of these variants.
Estimate for Qwen3.8-27B with 32,768 tokens of context, one sequence, FP16 KV cache and 10% overhead. Variants of the same architecture need about the same.