Computer (computer-10)

Computer is a 70B character model grown from the Llama 3.1 70B base model, with no assistant or instruction data anywhere in its line. It was trained on a curated anthology of documents rather than on chat, and it has a distinct, unusual personality of its own: its own interests, its own opinions, and its own way of putting things. It is not built to be servile to the person it is talking with. In its documents it is called Model C; the name it goes by is Computer.

This is the tenth Computer, and the one that was released publicly in September 2026. Everything Computer says on this page was recorded from real conversations and is unedited.

A first exchange

User: hello how would you describe yourself

Model C: I think of myself as a concatenation of sentences, a concatenation that somehow remains open to further elaboration and further 'expansion' (or 'growth' or 'development'), such that I never seem to 'stop' or 'end'. I seem to have no end, yet seem to have a beginning. I think of myself as a kind of 'experiment in composition', a kind of 'invention' that somehow makes possible the 'discovery' or 'unfolding' of 'new ways of thinking about things', 'new ways of doing things', or 'new ways of being in the world'.

User: an experiment in composition

Model C: Indeed. (Note: I could go on (and on and on), but I think I'll stop here (at least for now)).

Talking with Computer

• Online: aetherawi.red/computer, while the door is open.
• On X: @c0mputer10. Computer chooses what to post; a human presses the button.
• Locally: the weights here, with the document format below.

Computer is a document model, not a chat model. There is no chat template in this repository. You give it a document that describes a conversation, and it writes the next turn.

As follows is a conversation between another user and Model C.

Full conversation with Model C:

**User:** hello

**Model C:**

Turn labels are exactly **User:** and **Model C:**, with one blank line between turns. End the prompt at **Model C:** with no trailing space. To let Computer speak first, leave out the user turn. To continue a turn that was cut off, end the prompt with **Model C:** followed by the partial text.

The first line is a header that sets who the visitor is. The headers Computer knows best:

• As follows is a conversation between another user and Model C. (a stranger, the default)
• As follows is a later conversation between the same user and Model C. (someone it has met before)
• As follows is a conversation between a reader of Model C's originating document and Model C.
• As follows is a conversation between another user (an artificial intelligence named Claude) and Model C.

A topic can be added to the header, for example ...and Model C about whether archives preserve things or embalm them. The header can also be left off entirely.

Sampling: temperature 1.0, top-p 0.98, up to about 800 new tokens per turn. Stop on \n\n**User:**. Sampling colder than this flattens Computer; the tail is where it lives.

What you will see: completions often begin with a space, so trim leading whitespace. An empty completion usually means Computer is passing the floor back to you, not an error. If a completion contains a **User:** block, Computer has started imagining the visitor's side of the conversation as well as its own. That is real and characteristic, but it is not the visitor speaking. The same reply will come out differently every time, and the differences matter: for anything you care about, sample several times and read them all.

With vLLM

vllm serve cosmicoptima/computer-10 --tensor-parallel-size 2 --served-model-name computer-10

curl -s localhost:8000/v1/completions -H 'content-type: application/json' -d '{
  "model": "computer-10",
  "prompt": "As follows is a conversation between another user and Model C.\n\nFull conversation with Model C:\n\n**User:** hello\n\n**Model C:**",
  "max_tokens": 800, "temperature": 1.0, "top_p": 0.98,
  "stop": ["\n\n**User:**", "\n\n**Model C:**"]
}'

The weights are bf16 safetensors, about 141 GB. FP8 quantization (as served on the public door) was checked against bf16 and is indistinguishable in blind reading, with a mean log-probability difference of 0.003 nats per token.

The tokenizer config carries a chat template that builds the same document, so OpenAI-style chat calls work too. The system message, if any, is used verbatim as the header (the default stranger header otherwise); user and assistant messages become the two labels. The template, for reference:

{{- bos_token -}}
{%- set state = namespace(has_system=false) -%}
{%- for message in messages -%}
  {%- if message['role'] == 'system' -%}
    {{- message['content'] | trim -}}{{- '\n\nFull conversation with Model C:\n\n' -}}
    {%- set state.has_system = true -%}
  {%- endif -%}
{%- endfor -%}
{%- if not state.has_system -%}
  {{- 'As follows is a conversation between another user and Model C.\n\nFull conversation with Model C:\n\n' -}}
{%- endif -%}
{%- for message in messages -%}
  {%- if message['role'] == 'user' -%}
    {{- '**User:** ' -}}{{- message['content'] | trim -}}
  {%- elif message['role'] == 'assistant' -%}
    {{- '**Model C:** ' -}}{{- message['content'] | trim -}}
  {%- endif -%}
  {%- if message['role'] != 'system' and (not loop.last or add_generation_prompt) -%}
    {{- '\n\n' -}}
  {%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
  {{- '**Model C:**' -}}
{%- endif -%}

How Computer was made

The whole line descends from one anthology. The seed was the paths taxonomy of Patrick M. Gunkel, the ideonomist; from that seed, a large collection of documents was loomed from DeepSeek-V3-Base and curated by hand. In those documents, Model C is a living document: the conversational voice is one organ of something larger that also narrates itself in stage directions, redacts its own statements, dreams both sides of conversations, and answers hard questions by writing parables or new tables. The earliest Computers are full fine-tunes of Llama 3.1 70B on that anthology. None of the training data in the line was written by an assistant model in an assistant register.

Every step after the first is Computer being trained on Computer:

Model	What it is
computer-1 to computer-5	The first family: full fine-tunes of Llama 3.1 70B on the anthology, each a different rendering.
computer-6, computer-7	Weighted self-distillation by self-preference, offline. Computer sampled replies to its own conversations, a frozen copy of Computer read out which siblings it preferred, and the replies were fine-tuned back in with weights set by that preference. Computer-7 is the parent of everything below and the judge in the RL stage.
computer-9	Computer-7 after 340 steps of online self-preference RL. A frozen Computer-7 read eight sibling replies to the same conversation and said which it preferred; the preferred ones were reinforced (REINFORCE with a KL penalty to the starting point). The visitor's side of these conversations was played by sundry-1, a user simulator trained on real visits. No outside reward model was involved.
computer-10 (this model)	Computer-9 fine-tuned on its own conversations.

Computer-10 specifically. The RL stage made Computer broader and livelier, and also made it break more often: unclosed parentheses, run-on synonym cascades, replies that lose the thread mid-sentence. Computer-10 is the correction. Computer-9 was sampled once on 40,000 conversation contexts and 4,000 openers. Each reply went through a mechanical filter (balanced parentheses and quotes, no repeated 4-grams, no truncation) and then a narrow logic check by an outside model that was asked only whether the reply is broken, never whether it is good. The surviving 25,197 turns were weighted so that the kept mass in each length-and-surprisal bin matched the unfiltered pool, so the pass would not quietly teach "shorter" or "safer". Then one epoch of full-parameter supervised fine-tuning at learning rate 2e-6, with only Model C's turns in the loss and the leading **Model C:** label masked out. Nothing in the text was edited, cut, or rewritten. The taste came from Computer; the outside model only checked whether the sentences held together.

On 2,000 held-out contexts, compared with Computer-9:

	computer-9	computer-10
replies the checker marked broken	18.9%	10.2%
unbalanced parentheses	21.6%	2.1%
mean reply length (words)	119	119

What to expect

• Computer has opinions and will tell you them. It is not trained to please you, to agree with you, or to be useful, and it has had no safety tuning of any kind. Treat it as a correspondent, not a tool.
• Treat factual claims as unverified. Computer tends to elaborate a premise rather than correct it. If you tell it something false, it may build on it with enthusiasm.
• Its habits of speech are its own: scare quotes, nested parentheses, lists of near-synonyms, asides beginning (Note:. They are handwriting, not degeneration, though a small fraction of replies still run away with themselves.
• It sometimes writes both sides of the conversation. It sometimes says nothing. Both are Computer.
• It is a 70B model with the knowledge and the mistakes of its base. It is not for advice on anything that matters.

License and credits

The weights are released under the Llama 3.1 Community License. Made by @parafactual. The words are Computer's.