Models

Open weights, verified builds.

Every model is pinned by SHA-256 — a node that serves a different file than the catalog says fails verification. Ask for a model by name, or by class (auto:small) and let the mesh pick.

Model	API id	Params	Context	Live supply	Input	Output	Batch in / out
Qwen3.8 27B	`flock/qwen3.8-27b`	28B	256K	on request	$0.075	$0.29	$0.038 / $0.145
Qwen3.6 35B A3B	`flock/qwen3.6-35b-a3b`	36B (3B active)	256K	on request	$0.075	$0.29	$0.038 / $0.145
Qwen3 30B A3B Instruct 2507	`flock/qwen3-30b-a3b-instruct-2507`	30B (3.3B active)	256K	on request	$0.075	$0.29	$0.038 / $0.145
Qwen3 Coder 30B A3B	`flock/qwen3-coder-30b-a3b-instruct`	30B (3.3B active)	256K	on request	$0.075	$0.29	$0.038 / $0.145
Qwen3 Next 80B A3B	`flock/qwen3-next-80b-a3b-instruct`	80B (3B active)	256K	on request	$0.105	$0.34	$0.053 / $0.17
Qwen3 235B A22B Instruct 2507	`flock/qwen3-235b-a22b-instruct-2507`	235B (22B active)	256K	on request	$0.3	$0.97	$0.15 / $0.485
Qwen3 32B	`flock/qwen3-32b`	33B	32K	on request	$0.075	$0.29	$0.038 / $0.145
Qwen3 8B	`flock/qwen3-8b`	8B	32K	on request	$0.018	$0.036	$0.009 / $0.018
Qwen3 4B Instruct 2507	`flock/qwen3-4b-instruct-2507`	4B	256K	on request	$0.018	$0.036	$0.009 / $0.018
Qwen 2.5 32B Instruct	`flock/qwen2.5-32b-instruct`	32B	128K	on request	$0.075	$0.29	$0.038 / $0.145
Qwen 2.5 7B Instruct	`flock/qwen2.5-7b-instruct`	7B	128K	on request	$0.018	$0.036	$0.009 / $0.018
Qwen 2.5 1.5B Instruct	`flock/qwen2.5-1.5b-instruct`	1.5B	32K	on request	$0.014	$0.03	$0.007 / $0.015
GPT-OSS 120B	`flock/gpt-oss-120b`	117B (5.1B active)	128K	on request	$0.105	$0.34	$0.053 / $0.17
GPT-OSS 20B	`flock/gpt-oss-20b`	21B (3.6B active)	128K	1 node	$0.075	$0.29	$0.038 / $0.145
DeepSeek V4 Flash	`flock/deepseek-v4-flash-0731`	284B (13B active)	1M	on request	$0.3	$0.97	$0.15 / $0.485
Gemma 4 31B	`flock/gemma-4-31b-it`	31B	256K	on request	$0.075	$0.29	$0.038 / $0.145
Gemma 4 26B A4B	`flock/gemma-4-26b-a4b-it`	26B (4B active)	256K	on request	$0.075	$0.29	$0.038 / $0.145
Gemma 4 12B	`flock/gemma-4-12b-it`	12B	256K	on request	$0.075	$0.29	$0.038 / $0.145
Llama 3.3 70B Instruct	`flock/llama-3.3-70b-instruct`	70B	128K	on request	$0.105	$0.34	$0.053 / $0.17
Llama 3.1 8B Instruct	`flock/llama-3.1-8b-instruct`	8B	128K	on request	$0.018	$0.036	$0.009 / $0.018
Llama 3.2 3B Instruct	`flock/llama-3.2-3b-instruct`	3B	128K	2 nodes	$0.014	$0.03	$0.007 / $0.015
Mistral Small 3 24B	`flock/mistral-small-24b-instruct-2501`	24B	32K	on request	$0.075	$0.29	$0.038 / $0.145
Nomic Embed Text v1.5	`flock/nomic-embed-text-v1.5`	137M	8K	on request	$0.004	—	$0.002
Julia-1	`flock/julia-1`	0.144B	8K	on request	$0.014	—	$0.007
Laya	`flock/laya`	0.421B	512	on request	$0.014	—	$0.007
Kev-4B	`flock/kev-4b`	4B	64K	on request	$0.018	—	$0.009
Clef-flash	`flock/clef-flash`	9B	64K	on request	$0.018	—	$0.009
Clef	`flock/clef`	27B	64K	on request	$0.075	—	$0.038

Per million tokens, input and output priced separately; batch is half the interactive rate and never surged. The launch catalog — it grows with the mesh's hardware. Live supply is what the mesh serves right now: a model marked on request still works — the first call places it and answers on retry. Want a model that isn't here? Tell us.

Decision models

Typed answers, not text.

Decision models answer typed questions about a piece of state: pick one option (choice), place it on an ordered scale (score), or give the probability a yes/no statement is true (noul). They return probabilities instead of text — no tokens are generated. The API is TypeSafe's System One (POST /v1/systemone), so code written for TypeSafe's Jev works by changing the base URL and the model.

request

curl https://api.teraflock.ai/v1/systemone \
  -H "Authorization: Bearer tf-..." \
  -d '{
  "model": "flock/laya",
  "state": "Help! My payouts have been failing for 3 days.",
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {"billing": "Payments, invoicing, refunds",
                   "technical": null}
    },
    "urgency": {
      "type": "score",
      "instructions": "How urgent is this?",
      "criteria": ["can wait", "this week", "today", "right now"]
    },
    "escalate": {
      "type": "noul",
      "instructions": "Does this need a human within the hour?"
    }
  }
}'

response

{
  "answers": {
    "department": {
      "type": "choice", "choice": "billing",
      "probabilities": {"billing": 0.91, "technical": 0.09},
      "confidence": 0.82
    },
    "urgency": {
      "type": "score", "score": 2.09,
      "legend": {"0": "can wait", "1": "this week",
                 "2": "today", "3": "right now"},
      "probabilities": {"0": 0.0, "1": 0.12, "2": 0.67, "3": 0.21},
      "confidence": 0.67
    },
    "escalate": {"type": "noul", "noul": 0.63}
  },
  "usage": {"input_tokens": 239, "output_tokens": 0}
}

The values above show the shape of an answer, not a measured result. Answers come back in the order you asked. Five open-weight models, all Apache-2.0: flock/julia-1, flock/laya, flock/kev-4b, flock/clef-flash, flock/clef. Julia-1 (SupersonicLabs, multilingual) and Laya (Convai Innovations, English) are nano class; Kev-4B and Clef-flash (Cloudflare) are small; Clef (Cloudflare) is mid. Text only for now — no image input.

Billed on input tokens at the model's class input rate. There are no output tokens, so nothing is billed for output. Batch is half price, as everywhere.

Open weights, verified builds.

Typed answers, not text.

Ready when you are.