Abliterated Model Large V2

Abliterated Model Large V2 is a hosted, text-only reasoning large language model offered by Abliteration.ai under the API identifier abliterated-model-large-v2. Abliteration.ai announced the service on August 31, 2026 and describes it as a derivative of GLM-5.3 that the company further modified to reduce refusal behavior and fine-tuned for its platform.^[1][2] The provider markets it as an "uncensored" model for harder reasoning, evaluation, red teaming, agent testing, and offensive cybersecurity work.^[1][2] Those descriptions are provider claims, not results from an independent evaluation of this derivative.

The service has a one-million-token combined context window, accepts only text, and exposes low, high, and max reasoning modes.^[2][3] Abliteration.ai lists a maximum output of 999,990 tokens when the prompt is small enough, but it does not publish a tested output-length distribution or a reliability study for responses near that limit.^[2] As of September 2, 2026, the provider had not linked public V2 weights, a model repository, an artifact license, an exact GLM-5.3 checkpoint or base revision, a reproducible modification procedure, a model card, a system card, or an evaluation performed on the modified service.^{[2][9][11][12]}

Scroll sideways for more →

Release and model identity

Abliteration.ai positions Large V2 as its default large model and labels the earlier abliterated-model-large service its previous large model. Its documentation distinguishes the versions by their provider-described bases: Large V2 is associated with GLM-5.3, while the previous large service is associated with GLM-5.2.^[2] The documentation also says that Abliteration.ai both abliterated and fine-tuned Large V2. It does not identify the fine-tuning data, objective, number of training steps, parameter changes, or the exact upstream checkpoint.

The release announcement says the intended workloads include offensive cybersecurity, red-team exercises, and agent evaluation.^[1] This describes the provider's positioning. It does not establish measured performance on those workloads.

The announcement also refers to GLM-5.3 benchmark rankings and to Z.ai's comparison of GLM-5.3 with GLM-5.2 on cyber exploitation tests.^[1] Z.ai publishes weights and evaluation material for its own GLM-5.3 checkpoint.^[9] Abliteration.ai did not publish a corresponding rerun for abliterated-model-large-v2, so the base model's scores, safety behavior, and capability profile cannot be assigned to the modified hosted service.

Hosted API behavior

The provider exposes Large V2 through OpenAI-compatible Chat Completions and Responses endpoints and an Anthropic-compatible Messages endpoint. Its compatibility table lists streaming, function calling, structured output, reasoning controls, and server-side web search across the API surfaces.^[4] These are documented service interfaces. The public materials do not include conformance tests, tool-call accuracy measurements, or a complete record of differences from the APIs they emulate.

Large V2 is text-only. Abliteration.ai says that image or video content sent to the model is rejected with an HTTP 400 response, even though its general-purpose abliterated-model service accepts media on some endpoints.^[2][4] This distinction matters because some provider-wide feature tables include image and video rows that do not apply to Large V2.

Reasoning is enabled by default at max effort. The API accepts several compatibility labels, but the provider maps them to three actual Large V2 modes: none, minimal, and low map to low; medium and high map to high; xhigh and max map to max.^[3] A request to disable reasoning does not turn it off. The service instead runs low-effort reasoning and hides the trace. Hiding a trace does not remove its reasoning-token cost.^[3]

The one-million-token limit covers the prompt and completion together. The documented 999,990-token maximum output is therefore available only with a very small prompt.^[2] Neither the model documentation nor the release post provides long-context retrieval results, latency at different lengths, memory details, or a quality evaluation near the stated limit.

Pricing and data handling

Abliteration.ai's rate table listed the same price for uncached input and output on September 2, 2026: $5 per million tokens for each. Recognized prompt caching reads cost $0.50 per million input tokens. Cache creation is charged at the ordinary input rate, and the service reports cached and uncached input separately in the usage record.^[5] These are mutable hosted-service prices rather than properties of the underlying model.

The company says prompts and completions on its core inference endpoints are processed in memory, are not written to disk, and are not used for training. It retains operational telemetry such as token counts, timestamps, model identifiers, HTTP status codes, and billing records.^[6] The same policy says that optional web search and web fetch send a query or target URL to an outside provider, so the company's zero-retention description does not apply to those tool calls. Its training-data generator also stores generated datasets so that customers can review and export them.^[6]

These statements describe Abliteration.ai's policy. The reviewed sources do not include an independent retention audit, an infrastructure assessment, or enough implementation detail to verify the policy's enforcement. "Zero data retention" should therefore be read as the provider's declared handling practice, not as an independently certified property of Large V2.

Abliteration and refusal removal

In model research, abliteration commonly refers to changing model weights to suppress a direction associated with refusal behavior. Arditi and coauthors identified a one-dimensional refusal-related subspace across 13 open-source chat models up to 72 billion parameters. They showed that removing the direction from residual-stream activations or weight matrices could reduce refusals in those models, while adding it could induce refusals on harmless prompts.^[7] Their study predates GLM-5.3 and does not test Abliteration.ai's service.

Abliteration.ai says that its abliteration process removes a refusal direction at the weight level and that Large V2 was also fine-tuned after its GLM-5.3 base was modified.^[2][10] Without V2 weights, activation measurements, code, a modification recipe, or the fine-tuning record, an outside reviewer cannot reproduce that process or determine how closely it follows the method studied by Arditi and coauthors.

Refusal removal also does not prove that other behavior is unchanged. A 2026 preprint by Aleksander Fafuła compared base and abliterated checkpoints from two other mixture-of-experts model families on 21,600 decisions under uncertainty. The abliterated variants made more optimistic choices, produced longer justifications, and used fewer explicit uncertainty terms, while expressed confidence moved in opposite directions across the two families.^[8] That study does not evaluate GLM-5.3 or Large V2. It shows why behavior measured on a base model, or on another abliterated model, should not be treated as evidence for this service.

Evidence limits

Large V2 has a stable model identifier, current provider documentation, published API limits, and a public price. Those records establish that it is a named hosted service rather than an inferred or rumored model. They do not make its modified weights or development process inspectable.

Scroll sideways for more →

Until derivative-specific evidence is published, GLM-5.3 results can describe only the provider-identified starting point. They cannot establish Large V2's accuracy, refusal rate, safety profile, cyber capability, long-context reliability, or performance relative to other models.

References

1. ^Abliteration.ai (@abliteration_ai). "Abliterated Model Large V2 release announcement." X, August 31, 2026. x.com/...2094458081451393287
2. ^Abliteration.ai. "abliteration.ai models." Last modified August 31, 2026. docs.abliteration.ai/models
3. ^Abliteration.ai. "Thinking and reasoning effort." Last modified August 31, 2026. docs.abliteration.ai/...thinking
4. ^Abliteration.ai. "Compatibility matrix." Last modified August 31, 2026. docs.abliteration.ai/compatibility-matrix
5. ^Abliteration.ai. "Pricing for LLM API, Credits, and Policy Gateway." Accessed September 2, 2026. abliteration.ai/pricing
6. ^Abliteration.ai. "Data controls at abliteration.ai." Last updated May 29, 2026. abliteration.ai/data-handling
7. ^Arditi, Andy, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. "Refusal in Language Models Is Mediated by a Single Direction." arXiv:2406.11717, revised October 30, 2024. arxiv.org/...2406.11717
8. ^Fafuła, Aleksander. "Abliteration Is Not a Scalpel: Off-Target Effects of Refusal Removal on Decision Disposition Across Model Families." arXiv:2607.17427, July 19, 2026. arxiv.org/...2607.17427
9. ^Z.ai. "GLM-5.3" model card and repository. Hugging Face, accessed September 2, 2026. huggingface.co/...GLM-5.3
10. ^Abliteration.ai. "What is abliteration?" Last modified August 30, 2026. docs.abliteration.ai/what-is-abliteration
11. ^Abliteration.ai organization. Model listing showing no public models. Hugging Face, accessed September 2, 2026. huggingface.co/abliterationaiorg
12. ^Abliteration AI organization. Repository listing. GitHub, accessed September 2, 2026. github.com/abliteration-ai