All prices are per million tokens. Checkpoint storage is charged at $0.10 per GB per month.
Training #
We provide an 80% discount on cached prefill tokens.
| Model | Tinker ID | Context | Size | Arch | Type | PrefillCached: 80% discount | Sample | Train |
|---|---|---|---|---|---|---|---|---|
Inkling (256K)Limited-time 50% discountInkling-SmallLimited-time 50% discountInkling-Small (256K)Limited-time 50% discountNemotron-3.5-Lightning-30B-A3BLimited-time 50% discountNemotron-3.5-Lightning-30B-A3B (256K)Limited-time 50% discountNemotron-3-Ultra-550B-A55BLimited-time 50% discountNemotron-3-Ultra-550B-A55B (256K)Limited-time 50% discountNemotron-3-Super-120B-A12BLimited-time 50% discountNemotron-3-Super-120B-A12B (256K)Limited-time 50% discountNemotron-3-Nano-30B-A3BLimited-time 50% discountGLM-5.3 (256K)Kimi-K2.6Kimi-K2.6 (128K)Qwen3.8-27BQwen3.8-27B (256K)Qwen3.6-35B-A3BQwen3.6-27BRetiring September 2Qwen3.5-397B-A17BQwen3.5-397B-A17B (256K)Qwen3.5-35B-A3B-BaseQwen3.5-9BQwen3.5-9B-BaseQwen3.5-4BQwen3-8BGPT-OSS-120BGPT-OSS-120B (128K)GPT-OSS-20BDeepSeek-V3.1## Serverless Inference (Beta)
Serverless inference is currently in beta and available for Inkling and Inkling-Small only. We do not recommend it for intensive production use until it is out of beta. If you're interested in production use, email us at [email protected] to join the waitlist. Please include which models you need, your expected volume and latency requirements, and your use case.
| Model | Tinker ID | Context | Prefill (Input) | Sample (Output) |
|---|---|---|---|---|
Inkling## Pricing Terms
Prefill: Processing input/prompt tokens (forward pass only)** Cached prefill**: The smaller price under each prefill price; applies to input tokens that hit the prompt cache (80% off)** Sample**: Generating output tokens (forward pass + sampling)** Train**: Forward and backward pass for gradient computation** Context**: Maximum sequence length. Models with:peft:
suffix support extended context at higher prices.Tinker ID: The exact string to pass tocreate_lora_training_client(base_model=...)
orcreate_sampling_client(base_model=...)
MoE models are priced by active parameters, making them significantly more cost-effective than dense models of similar quality.
Model Types #
Base: Raw pretrained models with no chat or instruction tuning. Best for post-training research or running the full post-training pipeline yourself.Reasoning: Always produce chain-of-thought before their answer. Highest intelligence, higher latency and token cost.** Hybrid**: Run in both thinking and non-thinking modes. They reason by default, but chain-of-thought can be disabled via a renderer or argument for faster, cheaper direct answers.Vision: Vision-language models that accept images alongside text. Shown as a+ Vision
suffix on the underlying type (for example,Hybrid + Vision
).Audio: Models that accept audio alongside text. Shown as a+ Audio
suffix on the underlying type.
Architecture is either Dense (all parameters active per token) or MoE (mixture-of-experts, only a subset of parameters active per token). MoE models are highlighted in amber.
Choosing a Model #
Cost-effective: Use MoE models (highlighted in amber)** Research/post-training**: Use Base models** Task-specific fine-tuning**: Start with a Hybrid model** Low latency**: Use a Hybrid model with chain-of-thought disabled** High intelligence**: Use Reasoning or Hybrid models (chain-of-thought)** Vision tasks**: Use models with Vision in the type
Machine-Readable Pricing #
If you want to use this data programmatically (cost estimation, model pickers, dashboards), don't scrape the tables above. The data behind them is published as JSON alongside this page, and those files are the stable interface for scripts (the tables' HTML is presentational and may change):
models.json: the Training table. One object per model withname
,tinker_id
,context
,size
,arch
,type
,url
, and per-million-token pricesprefill
,cached_prefill
,sample
,train
(strings like"$0.374"
). Temporarily discounted models also carryoriginal_*
price fields and anote
.serverless.json: the Serverless Inference table, withname
,tinker_id
,context
,url
,input
,cached_input
,output
.
For example:
import json, urllib.request
url = "https://tinker-docs.thinkingmachines.ai/tinker/models.json"
models = json.load(urllib.request.urlopen(url))
for m in models:
print(m["tinker_id"], m["train"])
Retired Models #
These models have been retired and can no longer be used for training or inference, grouped by retirement date. See Model deprecations for the recommended replacement for each.
July 12, 2026
Kimi:Kimi-K2.5
June 12, 2026
Qwen:Qwen3-235B-A22B-Instruct-2507
,Qwen3-VL-235B-A22B-Instruct
,Qwen3.5-35B-A3B
,Qwen3.5-27B
,Qwen3-32B
,Qwen3-30B-A3B
,Qwen3-30B-A3B-Instruct-2507
,Qwen3-VL-30B-A3B-Instruct
,Qwen3-30B-A3B-Base
,Qwen3-8B-Base
,Qwen3-4B-Instruct-2507
Llama:Llama-3.3-70B-Instruct
,Llama-3.1-70B
,Llama-3.1-8B
,Llama-3.1-8B-Instruct
,Llama-3.2-3B
,Llama-3.2-1B
DeepSeek:DeepSeek-V3.1-Base
Kimi:Kimi-K2-Thinking