Abliterated DeepSeek and GLM: OpenBlit model preview
Preview DeepSeek V4.1 Flash 1017 and GLM 5.3, compare published input/output prices, and understand OpenBlit’s API and launch-qualification boundaries.
OpenBlit is coming soon. Public accounts and inference are not open.
The two models in the preview catalog
OpenBlit’s published preview lists abliterated variants of DeepSeek V4.1 Flash 1017 and GLM 5.3. Both are intended for technical research and coding-agent workflows. A model-family name alone does not identify a particular edited checkpoint, serving configuration, or measured level of performance.
| Model | Input | Output |
|---|---|---|
| DeepSeek V4.1 Flash 1017 | $0.20 | $0.85 |
| GLM 5.3 | $1.20 | $3.50 |
These are the website’s published preview prices, not a live checkout or a promise that capacity is available. Prices and context limits remain subject to launch qualification. Input and generated output are different token categories; the table is not a measured cost for a particular agent task.
Compare the checkpoint on your task, not just its label
The preview positions DeepSeek for reasoning and tool use, and GLM for code-analysis and reverse-engineering workflows. Those descriptions are product intent, not a published head-to-head benchmark. We are not claiming that either model is universally better.
For an evaluation, use the same codebase, prompts, tool definitions, and success criteria. Check whether findings are supported by the supplied evidence, whether tool calls are valid, and whether the model can revise a mistaken hypothesis. If the edited checkpoint changes, repeat the comparison instead of carrying over a result from a standard or differently quantized model.
An OpenAI-compatible interface for existing tools
The planned client interface follows Chat Completions: a model identifier, a messages array, and the usual generation and streaming parameters. This is intended to work with compatible SDKs and agents that let you configure a provider endpoint. Compatibility describes the interface, not endorsement by OpenAI or identical model behavior.
The preview identifiers are deepseek-v4.1-flash-1017-abliterated and glm-5.3-abliterated. The homepage has illustrative configurations for Pi, OpenCode, Claude Code Router, Hermes, TypeScript, and cURL. They are examples, not credentials or confirmation of a working public inference endpoint.
At present, /v1 requests return an inference-unavailable error, and public login, signup, and console access are closed. No GPU inference is running on the website host.
What needs qualification before launch
The preview catalog shows a 1M-class context target. A long context listing is not evidence of effective recall, concurrency, or a response-time guarantee. Those depend on the exact model artifact, cache behavior, serving engine, workload, and hardware.
Before treating a configuration as production-ready, qualification needs to cover artifact identity, edited-model behavior, tool-use correctness, actual context handling, streaming and cancellation, and measured performance under the intended load. We do not transfer results from a different node or an unedited checkpoint to these product models.
OpenBlit’s privacy and crypto-payment descriptions express the intended product design. Payment crediting is not implemented, and this preview does not establish an audited retention policy or production service-level agreement.
Where to go next
Read what abliteration means if you are comparing an edited model with its original checkpoint. Use the security-research evaluation checklist to define a repeatable comparison instead of relying on a model’s label.
The homepage catalog remains the shared source for preview prices. When inference opens, check the actual service documentation and model availability before wiring an agent to a production workflow.