# Thinking Machines vs General Compute

> Thinking Machines sells Tinker for post-training open models. General Compute serves open models and fine-tuned checkpoints at high speed.

Canonical: https://www.subconscious.dev/compare/thinking-machines-vs-general-compute · By The Subconscious Team · Updated October 27, 2026

## How they compare

Thinking Machines' Tinker lets teams write their own SFT or RL loops over LoRA adapters on models like Kimi K2.6, GLM-5.3, DeepSeek-V3.1 and gpt-oss, billed per million tokens across prefill, sample and train. Its beta serverless API serves only the open Inkling models, at $1.00 in and $4.05 out for Inkling, and checkpoint sampling is scoped to testing. General Compute serves MiniMax M2.7, DeepSeek V3.2, DeepSeek V3.1 and GPT-OSS 120B from $0.21 in, and accepts LoRA or fully fine-tuned checkpoints behind a private model ID. It reports over 1,000 tokens per second on GPT-OSS 120B in its own benchmark.

Thinking Machines is the pick for custom post-training, including RL, without running GPU clusters, and for trying the Inkling models. General Compute fits production serving where speed matters. Both cover DeepSeek V3.1 and GPT-OSS, so a team could train on Tinker and serve the result on General Compute.

## What each one does

### Thinking Machines

Thinking Machines Lab is the San Francisco lab Mira Murati, formerly CTO of OpenAI, founded in February 2025, with OpenAI co-founder John Schulman as chief scientist. It raised about $2 billion at a $12 billion valuation in July 2025 in a round led by Andreessen Horowitz, and in March 2026 signed a multi-year Nvidia deal for one gigawatt of Vera Rubin capacity. In September 2026 it signed a $65M-a-year deal with Crusoe to serve Inkling, GLM 5.2 and 5.3 and its fine-tunes on dedicated B200s. Its main developer product is Tinker, an API for post-training open-weight models that launched in October 2025 and is now generally available. Tinker exposes four low-level calls, forward_backward, optim_step, sample and save_state, so teams write their own supervised or reinforcement learning loops while Thinking Machines runs the distributed GPU work. Training uses LoRA adapters rather than full weight updates.

### General Compute

General Compute is a San Francisco inference cloud built on alternative chips. It splits each request across hardware, with NVIDIA GPUs handling prefill and ASICs from Cerebras, SambaNova, Positron AI and d-Matrix handling decode. In its own open benchmark on GPT-OSS 120B it measured 1,025 to 1,234 output tokens per second against 100 to 121 on an unnamed GPU cloud, and up to 16.2x faster end to end. Founders Finn Puklowski and Jason Goodison raised a $15M seed led by FUSE, an up to $400M debt facility from Upper90 and, on October 27, 2026, a $100M Series A co-led by Blackbird and Abstract.

## Which is best, and when

### Choose Thinking Machines for

- Custom SFT and RL post-training
- Fine-tuning large MoE models
- The open Inkling models

### Choose General Compute for

- Fast production serving
- Hosting fine-tuned checkpoints
- Real-time agent loops

## At a glance

| Attribute | Thinking Machines | General Compute |
|---|---|---|
| Model access | Open weights | Open weights |
| Flagship models | Inkling, Inkling-Small | MiniMax M2.7, DeepSeek V3.2, GPT-OSS 120B |
| Speed | - | 1,000+ tok/s on GPT-OSS 120B (vendor) |
| Price | Per 1M tokens by prefill, sample, train; Inkling $1.00 in, $4.05 out | $0.21–$0.28 in, $0.38–$1.20 out per 1M |
| Customization | LoRA SFT and RL via Tinker | Custom LoRA and fine-tuned checkpoints |
| Deployment | Training API, beta serverless (Inkling only) | Serverless API, dedicated racks, bare metal |
| Long context | Inkling up to 1M (256K serverless); Tinker 32K–256K | 192K (MiniMax M2.7) |

## FAQ

### What is the difference between Thinking Machines and General Compute?

Thinking Machines sells Tinker for post-training open models. General Compute serves open models and fine-tuned checkpoints at high speed.

### When should I choose Thinking Machines over General Compute?

Custom SFT and RL post-training; Fine-tuning large MoE models; The open Inkling models.

### When should I choose General Compute over Thinking Machines?

Fast production serving; Hosting fine-tuned checkpoints; Real-time agent loops.

### Is Thinking Machines or General Compute cheaper?

Thinking Machines: Per 1M tokens by prefill, sample, train; Inkling $1.00 in, $4.05 out. General Compute: $0.21–$0.28 in, $0.38–$1.20 out per 1M. The cheaper choice depends on the model and workload.

### Which has more context, Thinking Machines or General Compute?

Thinking Machines: Inkling up to 1M (256K serverless); Tinker 32K–256K. General Compute: 192K (MiniMax M2.7).

## Running long-horizon agents?

If your agents run past 200K tokens, compare both against Subconscious: [Subconscious vs Thinking Machines](https://www.subconscious.dev/compare/subconscious-vs-thinking-machines.md), [Subconscious vs General Compute](https://www.subconscious.dev/compare/subconscious-vs-general-compute.md).

Full profiles: [Thinking Machines](https://www.subconscious.dev/providers/thinking-machines.md), [General Compute](https://www.subconscious.dev/providers/general-compute.md).