Subconscious chat completions

Subconscious is OpenAI-SDK-compatible inference on a fine-tuned Qwen-3.6 served via sglang. You point an OpenAI client at our base URL; it works with standard OpenAI fields plus one knob for reasoning.

The whole product, in 6 lines

from openai import OpenAI

client = OpenAI(
    base_url="https://api.subconscious.dev/v1",
    api_key="sky_...",  # or set SUBCONSCIOUS_API_KEY
)

resp = client.chat.completions.create(
    model="subconscious/tim-qwen3.6-27b",
    messages=[{"role": "user", "content": "Hello"}],
)

That's the entire surface. The rest of this doc is about doing it well.

Critical context — read first

1. One base URL. https://api.subconscious.dev/v1. Inference is served via sglang behind a Baseten gateway. Endpoints: POST /chat/completions (primary), POST /completions (legacy OpenAI text completions), GET /models. POST /v1/responses exists in URL-space but is not functional — returns sglang errors for every input shape.
2. One model. subconscious/tim-qwen3.6-27b — fine-tuned Qwen-3.6, fp8 quantization. Get its full capability spec from GET /v1/models (yes, that endpoint exists).
3. Context: millions of tokens per the public dashboard / TIMRUN positioning. The /v1/models endpoint currently reports contextlength: 8192 and maxcompletion_tokens: 5000 — treat that as the safe-bet runtime limit at this deployment, not the product ceiling. When the dashboard and the API response disagree, default to the dashboard's claim and verify on your workload.
4. Modalities are text-first: model spec says input_modalities: ["text"], but vision actually works in practice for many image sources (placeholders, data URLs). Wikipedia URLs consistently 500. Audio, file, and video blocks all return 400. See references/multimodal.md for full coverage — treat vision as undocumented-but-functional, not officially supported.
5. Thinking is ON by default. When on, the model prepends a "Here's a thinking process:" prose block to the answer (no <think> tags — just plain text). Opt out by setting chattemplatekwargs.enable_thinking: false. This is inverted from OpenAI / ChatGPT.
6. Only temperature and stop are honored as sampling parameters per the model spec. Other fields like topp, seed, n, frequencypenalty, logit_bias, etc. return 200 but Baseten / sglang silently drops them.

Config

# .env
SUBCONSCIOUS_API_KEY=sky_yNZq...

The OpenAI SDK reads OPENAIAPIKEY by default — pass apikey= explicitly, or set OPENAIAPI_KEY to your Subconscious key. Don't ship a real OpenAI key against our base URL.

Auth

Authorization: Bearer <SUBCONSCIOUS_API_KEY>

What the model honors

The dashboard markets TIM-Qwen3.6-27B with:

• Millions of tokens of context
• Frontier-grade agent performance
• 75% of base-model compute
• 50% faster sustained throughput

These are the product-level claims — treat them as canonical.

GET /v1/models returns the current runtime snapshot, which is narrower:

{
  "id": "subconscious/tim-qwen3.6-27b",
  "context_length": 8192,
  "max_completion_tokens": 5000,
  "quantization": "fp8",
  "supported_sampling_parameters": ["temperature", "stop"],
  "supported_features": ["tools", "json_mode", "structured_outputs", "reasoning"],
  "input_modalities": ["text"],
  "output_modalities": ["text"]
}

Standard OpenAI request fields that do affect output:

Accepted but silently dropped / probably ignored (verified to return 200; not in the model's supportedsamplingparameters):

• Sampling: topp, topk, seed, frequencypenalty, presencepenalty, logitbias, logprobs, toplogprobs, paralleltoolcalls
• Identity / caching: user, safetyidentifier, promptcachekey, promptcacheretention, servicetier (any of auto / default / flex / priority)
• Storage: store, metadata
• Output config: modalities (this field controls OUTPUT modalities — ["text", "audio"] etc. — and the model only outputs text, so audio output isn't produced; input vision via imageurl content blocks is a separate thing that DOES work — see references/multimodal.md), audio (audio-output voice/format config — no audio output), verbosity, prediction, websearch_options
• Deprecated: functioncall, functions (use toolchoice and tools instead)

If your app depends on any of these to actually affect behavior, that won't work on Subconscious.

Rejected with 400:

• Custom tool type ({type: "custom", ...}) — only function tools work
• developer role messages
• function role messages (deprecated)
• reasoningeffort=minimal and reasoningeffort=xhigh (the other 4 enum values pass schema validation)

The one Subconscious-specific knob: chattemplatekwargs.enable_thinking

Pass via the OpenAI SDK's extra_body:

extra_body={
    "chat_template_kwargs": {"enable_thinking": False},
}

• Default (on): model emits a meta-reasoning preamble starting with something like "Here's a thinking process:" followed by numbered steps, then the actual answer. No tags wrap it — it's all plain prose in message.content.
• enable_thinking: false: model skips the preamble and answers directly.

For most user-facing chat / structured-output cases you'll want false. See references/thinking.md.

streamoptions.includeusage: true is also useful when streaming — that field IS standard OpenAI, not Subconscious-specific.

Tools — standard OpenAI function calling

Only the standard OpenAI tool shape works:

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get current weather in a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

resp = client.chat.completions.create(
    model="subconscious/tim-qwen3.6-27b",
    messages=[{"role": "user", "content": "Weather in Boston?"}],
    tools=tools,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)

tool_calls = resp.choices[0].message.tool_calls
# [{ id: "call_...", function: { name: "get_weather", arguments: '{"city":"Boston"}' } }]

The model emits a tool_calls array; your code executes the function, then sends a follow-up message with role: "tool" containing the result. Standard OpenAI tool loop — Subconscious does not execute tools server-side here.

Subconscious-specific tool types (type: "platform", MCP, URL-based function tools) are not supported on this endpoint — they return 400 with a missing-function validation error.

Structured output

schema = {
    "type": "object",
    "properties": {
        "sentiment": {"type": "string", "enum": ["positive", "neutral", "negative"]},
        "confidence": {"type": "number"},
    },
    "required": ["sentiment", "confidence"],
}

resp = client.chat.completions.create(
    model="subconscious/tim-qwen3.6-27b",
    messages=[{"role": "user", "content": "Analyze: 'works great'"}],
    response_format={
        "type": "json_schema",
        "json_schema": {"name": "sentiment", "schema": schema},
    },
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
import json
result = json.loads(resp.choices[0].message.content)

jsonmode and structuredoutputs are both in supportedfeatures. When enablethinking: false, you get clean JSON; with thinking on, you'll have to strip the prose preamble before parsing.

Streaming

stream = client.chat.completions.create(
    model="subconscious/tim-qwen3.6-27b",
    messages=[{"role": "user", "content": "Haiku about Boston."}],
    stream=True,
    stream_options={"include_usage": True},
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)

for chunk in stream:
    delta = chunk.choices[0].delta if chunk.choices else None
    if delta and delta.content:
        print(delta.content, end="", flush=True)
    if chunk.usage:
        print(f"\n[in={chunk.usage.prompt_tokens} out={chunk.usage.completion_tokens}]")

The final chunk before [DONE] carries usage when include_usage: true is set. Standard OpenAI SSE format.

Status codes

Error body shapes:

// sglang validation error
{"error": {"code": 400, "message": "...", "object": "error", "param": null, "type": "Bad Request"}}

// Simpler errors (auth, missing model)
{"error": "string"}

The OpenAI SDK raises typed exceptions (AuthenticationError, RateLimitError, BadRequestError, etc.) — prefer those over parsing the body.

Best practices

• Default enable_thinking: false for chat / structured / classification. Cuts latency dramatically; cleaner output.
• Leave enable_thinking: true for hard reasoning (research synthesis, code generation, planning).
• Set max_tokens on every call. No hard server default; runaway generations cost money.
• Use stream=True for any user-facing surface > 1s perceived latency.
• Add streamoptions.includeusage: true on streamed calls to track cost.
• Don't depend on topp, seed, frequencypenalty, etc. They return 200 but aren't honored.
• Bound message history reasonably. Dashboard claims millions of tokens; /v1/models reports 8192 at this deployment. Test the size you actually need — long prompts can silently truncate.
• Vision works but is undocumented. Data URLs (data:image/png;base64,...) are the reliable path; Wikipedia URLs always 500. Audio/file/video definitely don't work (400/500). See references/multimodal.md.

TypeScript / Node

import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://api.subconscious.dev/v1',
  apiKey: process.env.SUBCONSCIOUS_API_KEY!,
});

const resp = await client.chat.completions.create({
  model: 'subconscious/tim-qwen3.6-27b',
  messages: [{ role: 'user', content: 'Hello' }],
  // @ts-expect-error chat_template_kwargs is a Subconscious extension
  chat_template_kwargs: { enable_thinking: false },
});

cURL

curl https://api.subconscious.dev/v1/chat/completions \
  -H "Authorization: Bearer $SUBCONSCIOUS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "subconscious/tim-qwen3.6-27b",
    "messages": [{"role": "user", "content": "ping"}],
    "max_tokens": 50,
    "chat_template_kwargs": {"enable_thinking": false}
  }'

Pricing

Per the public dashboard at <https://www.subconscious.dev/platform>; (not verified against actual billing):

• Input: $0.50 / 1M tokens
• Output: $3.50 / 1M tokens

GET /v1/models returns zeros in its pricing field — that's a placeholder, not the rate. Real pricing comes from the dashboard / docs.

Thinking tokens count as output. Measured (median of 3 runs each, 2026-05-22):

For short prompts, thinking-on dominates cost. For long-output prompts, both modes can hit max_tokens.

Note: the pricing field returned by /v1/models currently shows zeros — that's the model spec endpoint, not the billing endpoint. Use the dashboard / docs for pricing.

References

• references/api-reference.md — Endpoint, full request/response shapes, error formats
• references/thinking.md — How thinking actually behaves; default-on; stripping the preamble
• references/tools.md — OpenAI function calling pattern; full tool loop; what's accepted vs rejected
• references/structured-output.md — response_format with JSON Schema + Pydantic/Zod
• references/multimodal.md — Vision (works on data URLs, sometimes URLs), audio/file/video (all rejected)
• references/best-practices.md — Production patterns, cost monitoring, retries
• references/examples.md — Working snippets (Python, TypeScript, cURL)

Resources

• API base: <https://api.subconscious.dev/v1>;
• Model spec: GET https://api.subconscious.dev/v1/models
• Dashboard: <https://www.subconscious.dev/platform>;
• Playground: <https://www.subconscious.dev/playground>;
• Docs: <https://docs.subconscious.dev>;
• Pricing: <https://docs.subconscious.dev/pricing>;