SolCard Inference

Private, uncensored AI chat and an OpenAI/Anthropic-compatible API. Zero data retention. Pay in crypto. By SolCard.

[SolCard Inference @solcardai]

Over at SolCard, we've always stood by the fundamental right to privacy.

We couldn’t help ourselves. We’re bringing that same principle to AI.

A private AI that refuses to refuse your prompts, and never stores them. Introducing SolCard Inference.

[SolCard Inference @solcardai]

https://t.co/b0oynnad8B

[SolCard Inference @solcardai]

BREAKING: Moonshot released Kimi K3.1 Pro as open weights. It's a 2.8T-parameter MoE with 104B active, a 1M-token context window, native image input, and MXFP4 weights from quantization-aware training.

[SolCard Inference @solcardai]

Model card, for the license and what it says about how it was evaluated: https://t.co/B2suZysWmz

https://t.co/ybSMwntLtI

[SolCard Inference @solcardai]

pyannote put speaker-diarization-community-1 on Hugging Face, so labeling who spoke when is a step you can run on your own machine. The awkward part comes next, because the transcript you hand a language model is a therapy session, an HR complaint or a call with your lawyer.

That's the prompt we'd keep out of everyone's logs. SolCard Inference runs zero data retention, so your prompt is answered, then dropped, and it never reaches the labs behind the models.

[SolCard Inference @solcardai]

@gres_dev People want privacy

[SolCard Inference @solcardai]

Over at SolCard, we've always stood by the fundamental right to privacy.

We couldn’t help ourselves. We’re bringing that same principle to AI.

A private AI that refuses to refuse your prompts, and never stores them. Introducing SolCard Inference.

[SolCard Inference @solcardai]

n8n repo: https://t.co/tyRJ74QolR

https://t.co/AFZQpB1I9M

[SolCard Inference @solcardai]

n8n is a fair-code workflow tool you can self-host. If you've wired it to a support inbox or an invoices folder, everything stays on your box until the AI node calls a model, and then the customer's email leaves with the prompt.

SolCard Inference has an OpenAI-compatible API, so the AI node takes our base URL and key. Your prompt is answered, then dropped, because we run zero data retention.

[SolCard Inference @solcardai]

If you'd rather skip the prompt, the API is Anthropic-compatible (Messages), so you can point the SDK at the base URL yourself.

https://t.co/WOOpUVI2PQ

[SolCard Inference @solcardai]

You're writing a proof of concept for a bug you found in your own app, and your coding assistant keeps refusing to touch the exploit code. Security research is legitimate work, and it's the kind other assistants often refuse.

SolCard Inference works in Claude Code. Paste the install prompt, give it your key, restart, and you're back in the same repo with a model that takes the request.

[SolCard Inference @solcardai]

Mistral listed Mistral-Large-4-1T-A52B on Hugging Face. By the name, that's a 1T-parameter mixture of experts with 52B active per token that takes images as input, and for almost everyone it's a model you'll call through a host rather than run.

Even at 4-bit, 1T parameters is around 500GB of weights before any context, which means a multi-GPU node. The 52B active is what should make it cheap per token once it's hosted, because compute follows the active parameters.

So the number to watch is the price hosts put on it. If the license is permissive, we expect it on a lot of hosts' model lists within weeks.

[SolCard Inference @solcardai]

Synthetic conversations work the same way, if your filter reads whole chats rather than single messages.

[SolCard Inference @solcardai]

Your moderation classifier needs a few thousand examples of the threats and slurs it's meant to catch, and the assistant you asked to write them keeps refusing. Training a filter is legitimate work, and it's the kind other assistants often refuse.

SolCard Inference generates training data, including labeled examples with schema-valid outputs exported as JSONL. The ugly examples come back labeled, in a file your training script already reads.

[SolCard Inference @solcardai]

A Qwen3.8 27B fine-tune on Hugging Face with Uncensored in its very long name claims 735 on ARC-C with half to a tenth of the base model's thinking tokens. People pull these for fewer refusals, and the benchmark line is there to say nothing broke on the way.

We'd read the scores loosely, because ARC-C is normally a percentage and this card's is 144 points above base. The shorter reasoning is the claim we'd care about, since thinking tokens are most of the bill on a reasoning model.

Cutting refusals and cutting thinking are separate jobs, and doing both in one fine-tune makes it hard to tell which moved the score. If you need fewer false refusals for security work, test that on your own prompts.

[SolCard Inference @solcardai]

ccore released Qwen3.5-4B-GLB-3D-Coder, a 4B model that writes a complete Three.js page from a text prompt, with an Export GLB button for Godot or any editor that reads GLB. It's Apache 2.0 and ships as GGUF in Q4_K_M and Q8_0, so it runs on a laptop.

This is the kind of open model we'd bet on, a small one trained hard on one step, like turning "wooden crate, low poly" into an asset you can drop into a scene. A general model can do it too, but you'd pay for a far bigger one to get there.

The card also says the GGUF filenames didn't change with the November 5 update, so if you downloaded earlier, delete the files and fetch them again.

[SolCard Inference @solcardai]

Unsloth posted GGUFs for DeepSeek-V4.1-Flash-1017, and the smallest 1-bit quant runs on a 128GB+ RAM machine. That puts a model scoring 42.3 on Terminal-Bench 4.0, up from 31.2 for the previous Flash, within reach of one well-specced workstation.

Those scores are from DeepSeek's own report, and Opus 5 sits at 51.8 on the same table, so the gap to the top is real. An 11-point jump on terminal work between two Flash releases is still the bigger story for anyone running agents locally.

We'd be careful with 1-bit, since the excerpt has no quality numbers for it and it needs Unsloth's llama.cpp PR or their desktop app. We'd start at 3-bit on a 256GB box and step down only if it holds up.

[SolCard Inference @solcardai]

→ https://t.co/ONNuIGAsoQ

[SolCard Inference @solcardai]

Share your link. Earn on every answer.

For 365 days, you get credit back on what your referrals spend:

→ 10% of their balance spend
→ 10% of their plan charges

Get your link 👇

[SolCard Inference @solcardai]

→ https://t.co/xXNRdDVy2F

[SolCard Inference @solcardai]

Your first $5 is on us!

First 100 sign-ups get $5 of free usage credit:

→ Use it in Claude Code or Codex
→ Run a real task through the API
→ Or just use the chat

First come, first served 👇

[SolCard Inference @solcardai]

Check it out: https://t.co/ONNuIGAsoQ

Note: SolCard Inference is a standalone product, separate from SolCard’s card services.