Thinking Machines Provider Guide

Thinking Machines is a Tier-2 catalog provider: OpenAI-wire-compatible with no behavioural quirks, so its entire integration is one JSON file (src/lib/providers/catalog/thinking-machines.json) rather than hand-written code. That file is the source of truth for everything on this page.

Verification status: this entry is docs-verified only, not yet live-verified. The vendor's GET /models needs a key (an unauthenticated GET answered HTTP 401), so the model ids come from the vendor's public Models & Pricing page and the roster has not been checked. No API key was created or used to build this entry. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

---

Key Facts

• Provider id: thinking-machines (alias tinker)
• Protocol: OpenAI-compatible (/chat/completions). The vendor's OpenAI-compatible page is titled "OpenAI API Compatible Inference (in beta)" and says "We support both /completions and /chat/completions endpoints."
• Base URL: https://tinker.thinkingmachines.dev/services/tinker-prod/oai/api/v1
• Default model: moonshotai/Kimi-K2.6
• Models in catalog: 12 rows of the Training table on the Models & Pricing page, listed under Models
• Vendor scope statement: "Currently, OpenAI-compatible inference is meant for testing and internal use with low internal traffic, rather than large, high-throughput, user-facing deployments." (OpenAI-compatible page)
• Streaming: true — the OpenAI-compatible page says "In streaming mode, reasoning_content and content arrive on separate SSE events"
• Thinking: true — the OpenAI-compatible page documents a separate_reasoning flag and a reasoning_effort parameter on /chat/completions, and says of the latter "Not all models support this parameter."
• Tool calling: not declared (false). With false, the provider's supportsTools() answers false, so no tools array is sent
• Structured output: not declared (false)
• Structured output + tools together: not declared (false) — no POST request was sent
• Embeddings: not declared (false)
• Billing: the catalog field holds no-free-tier, a schema placeholder (the schema has no unknown value), not a vendor statement; see Billing for what the vendor says
• Key format: none declared, so NeuroLink applies no key-format check

---

Quick Start

1. Get an API key

1. Visit: https://tinker.thinkingmachines.ai/keys — the Quickstart says "Or create an API key in the Tinker Console and set it as the TINKER_API_KEY environment variable." An unauthenticated GET of that URL redirects to a page titled "Sign in" (sign-in was not attempted while building this entry)
2. Authentication as the vendor states it: the OpenAI-compatible page says "Authenticate with your Tinker API key, by passing the same key used for Tinker as the API key to the OpenAI client." Its code example reads the key from TINKER_API_KEY
3. Billing as the vendor states it: see Billing
4. Set THINKING_MACHINES_API_KEY in your .env file (the vendor's examples read TINKER_API_KEY, which NeuroLink accepts as a fallback)

2. Configure

export THINKING_MACHINES_API_KEY=your-api-key
export THINKING_MACHINES_MODEL=moonshotai/Kimi-K2.6   # optional — overrides the default model
export THINKING_MACHINES_BASE_URL=https://tinker.thinkingmachines.dev/services/tinker-prod/oai/api/v1   # optional — proxy or self-hosted gateway

3. Use it

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "thinking-machines",
  model: "moonshotai/Kimi-K2.6",
});

console.log(result.content);

# CLI
npx @juspay/neurolink generate "Hello" --provider thinking-machines

Per-request credentials work as they do for every provider:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "thinking-machines",
  credentials: {
    thinkingMachines: { apiKey: process.env.THINKING_MACHINES_API_KEY },
  },
});

Sampler checkpoints

The OpenAI-compatible page says "Use a Tinker sampler weight path as the model name." and shows a path that begins with tinker://. A tinker:// path passed as model is an id outside the catalog, so it uses models.defaultContextWindow (32768) and models.defaultMaxOutputTokens (8192).

await neurolink.generate({
  input: { text: "Hello" },
  provider: "thinking-machines",
  model: "tinker://<your-sampler-checkpoint-path>",
});

---

Billing

The vendor's Quickstart says: "Set up payment info in Billing, and see Models & Pricing for current training and sampling rates." The Billing link goes to https://tinker.thinkingmachines.ai/billing/; an unauthenticated GET of it redirects to the same "Sign in" page as the keys URL. The Models & Pricing page says "All prices are per million tokens." The catalog's billingPolicy field holds no-free-tier only because the schema requires one of three values; it is not a vendor statement.

---

Models

Model	Context	Vision	$/M in · out (cached in)	Vendor type	Notes
moonshotai/Kimi-K2.6 ⭐	32,768	not declared	$2.205 / $5.49 ($0.441)	Hybrid + Vision	NeuroLink default. Named in the OpenCode tutorial's "Using a base model (no fine-tuning)" section.
Qwen/Qwen3.8-27B	65,536	not declared	$1.86 / $5.595 ($0.372)	Hybrid + Vision	Fallback. Listed under "Recommended replacement" for Qwen3.6-27B.
Qwen/Qwen3.6-35B-A3B	65,536	not declared	$0.54 / $1.335 ($0.108)	Hybrid + Vision	Listed under "Recommended replacement" for Qwen3.5-35B-A3B and Qwen3-30B-A3B.
Qwen/Qwen3.5-397B-A17B	65,536	not declared	$3.00 / $7.50 ($0.60)	Hybrid + Vision	Listed under "Recommended replacement" for Qwen3-235B-A22B-Instruct-2507 and Qwen3-VL-235B-A22B-Instruct.
Qwen/Qwen3.5-9B	65,536	not declared	$0.66 / $1.995 ($0.132)	Hybrid + Vision	Listed under "Recommended replacement" for Llama-3.1-8B-Instruct.
Qwen/Qwen3.5-4B	65,536	not declared	$0.33 / $1.005 ($0.066)	Hybrid + Vision	Listed under "Recommended replacement" for Qwen3-4B-Instruct-2507.
Qwen/Qwen3-8B	32,768	not declared	$0.195 / $0.60 ($0.039)	Hybrid	Fallback. Named on the Anthropic-Compatible API page, which documents a separate endpoint.
openai/gpt-oss-120b	32,768	not declared	$0.33 / $0.84 ($0.066)	Reasoning	Fallback.
openai/gpt-oss-20b	32,768	not declared	$0.18 / $0.45 ($0.036)	Reasoning	Listed in the Training table.
deepseek-ai/DeepSeek-V3.1	32,768	not declared	$1.695 / $4.215 ($0.339)	Hybrid	Listed in the Training table.
zai-org/GLM-5.3:peft:262144	262,144	not declared	$4.86 / $12.15 ($0.972)	Reasoning	The table row is named "GLM-5.3 (256K)".
nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16	65,536	not declared	$0.57 / $1.44 ($0.114)	Hybrid	Row note "Limited-time 50% discount"; the page shows Prefill $1.14 and Sample $2.88 struck through.

Context, prices and vendor type come from the Training table on the Models & Pricing page (retrieved 2026-09-29; the same data is published as models.json). Input is the Prefill column, output is the Sample column and the figure in brackets is the cached Prefill price; the page's Pricing Terms define Prefill as "Processing input/prompt tokens when sampling" and Sample as "Generating output tokens". Context is the table's label read with K = 1024 (the same table labels the id suffixes :peft:131072 and :peft:262144 as 128K and 256K).

models.defaultContextWindow (32768) is the figure in the OpenCode tutorial's Config fields table: "Max input tokens (32768 for most Tinker models)". models.defaultMaxOutputTokens (8192) is the "output" figure in both example configs of that tutorial, whose Config fields table describes limit.output as "Max output tokens". The catalog rows do not set maxOutputTokens, so the default applies to them.

Model ids: moonshotai/Kimi-K2.6 is the catalog id that the OpenCode tutorial shows as a models entry under its "Using a base model (no fine-tuning)" heading. The other eleven catalog ids are Tinker IDs from the Models & Pricing Training table. The verification level is docs-verified, not yet live-verified.

Vision: not declared means the entry sets vision to false: image input through /chat/completions is not declared by this entry. The Vendor type column shows the Training table's Type label, and the Model Types section describes Vision models as "Vision-language models that accept images alongside text."

Status: the rows' status is preview, following the OpenAI-compatible page's "(in beta)" title. The catalog schema requires a value; it is not a per-model vendor statement.

Fallback order when the default is unavailable: Qwen/Qwen3.8-27B → openai/gpt-oss-120b → Qwen/Qwen3-8B. The runtime fallback model name the loader derives (fallbacks[1]) is openai/gpt-oss-120b.

Retirements: the vendor's Model Deprecations page says "Avoid building hard dependencies on any specific model, as models may be updated, replaced, or removed over time."

---

Verification status

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog currently records for Thinking Machines — docs-verified only, not live-verified:

Probe	Result
Roster	unauthenticated GET /models answers HTTP 401 without a key (2026-09-29), so the twelve model ids are taken from the Models & Pricing page and the roster has not been checked
Endpoint shape	unauthenticated GET /chat/completions answered HTTP 405 with header allow: POST and body {"detail":"Method Not Allowed"}; the base URL itself answered HTTP 404 with body {"detail":"Not Found"}; no POST request was sent
Auth-failure shape	unauthenticated GET /models answered HTTP 401, content-type application/json, body {"detail":"X-Api-Key token is empty"}; the body has no code field, so no code is recorded and errorRules holds one rule, for status 401
Billing	the Quickstart's Billing section, quoted under Billing; no signup was performed
Tools / structured output	not declared; no POST request was sent — structuredOutputWithTools stays false
Live capability sweep	not run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=thinking-machines with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

---

Troubleshooting

Each row below is a statement from the vendor's own pages.

Symptom	What the vendor says	Source
HTTP 400 when a request sets reasoning_effort	"Not all models support this parameter. Requests against a model that doesn't support it return HTTP 400."	OpenAI-compatible page
Reasoning trace missing from content	"separate_reasoning defaults to true; set it to false to keep the reasoning trace inlined in content."	OpenAI-compatible page
Prompt formatting differs from the checkpoint	On /chat/completions "the server renders them with the model’s default Hugging Face chat template"; "If your checkpoint expects a different renderer, render the prompt to token IDs yourself"	OpenAI-compatible page
Latency or throughput varies	"Latency and throughput may vary by model and may change without notice during the beta." and "If you need higher or more stable throughput, contact the Tinker team in our Discord for guidance on larger-scale setups."	OpenAI-compatible page
A model id is no longer served	"Switch to the recommended replacements listed below for any deprecated models."	Model Deprecations page

---

See also

• Provider setup overview
• All providers
• Tier-2 onboarding — how this provider's JSON becomes a working integration
• Provider feature compatibility