<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/nVR2rRg0E" -->

---
title: Thinking Machines' v2 small model trained in 68 hours
description: Thinking Machines trained its v2 small model in 68 hours on GB300s, using NVFP4 across pre-, mid- and post-training, according to @soumithchintala....
canonical: https://daily.dev/posts/thinking-machines-v2-small-model-trained-in-68-hours-nvr2rrg0e
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Thinking Machines' v2 small model trained in 68 hours | daily.dev
og:description: Thinking Machines trained its v2 small model in 68 hours on GB300s, using NVFP4 across pre-, mid- and post-training, according to @soumithchintala....
og:url: https://daily.dev/posts/thinking-machines-v2-small-model-trained-in-68-hours-nvr2rrg0e
og:image: https://api.daily.dev/og/posts/nVR2rRg0E.png
og:image:alt: Thinking Machines' v2 small model trained in 68 hours
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Thinking Machines' v2 small model trained in 68 hours

**[Trends](https://daily.dev/sources/trends)** · 2 min read · 1 upvotes · 1 comments

## Summary

Thinking Machines trained its v2 small model in 68 hours on GB300s, using NVFP4 across pre-, mid- and post-training, according to @soumithchintala. Post-training added 17 million rollouts over 9 days. @cHHillee calls the infra 'really improved' and puts pretraining at just under 3 days. The model is live on Tinker now at $0.30 per million input tokens, $1.00 output and $1.50 for training. Open weights are promised next week, after external safety testing finishes. Vercel's @rauchg called the improvement rate 'bewildering' and says his team benchmarks it continuously.

## Content

The loudest reaction so far is about the training time, not the benchmarks. @yacineMTB: "This model trained so fast that genuinely, OpenAI or xai levels of compute could just hyper sweep it over a month." That's a pointed way of saying the real asset here is the pipeline.

The setup: @soumithchintala announced a v2 small model and says he's been using it daily for a variety of things. The numbers behind it are the part people keep repeating:

- Pre-training ran 68 hours on GB300s.
- Post-training was 17 million rollouts over 9 days.
- The whole thing, pre, mid and post training, was done in NVFP4.

It's available to train and fine-tune on Tinker now. Open weights come next week, once external safety testing is fully complete. Soumith's own parenthetical: "cant skimp on that." He frames the release as "continuing to decentralize and democratize intelligence."

@cHHillee, who's on the infra side, gave the clearest explanation of why it went so fast: "Our infra has really improved quite a bit since the last time we released a model." His version of the pretraining time is "just under 3 days," and in another telling "*barely* over 3 days." Either way, it's a tight turnaround for a model release.

He also pointed at price. On Tinker it's $0.30 for input, $1.00 for output and $1.50 for training, per the same post, and he hopes people find it "great for customization purposes." That positions it as a fine-tuning base more than a chat model to compete on leaderboards.

Outside the team, @rauchg called the rate of improvement "bewildering," saying it's now "even smaller, faster and more token-efficient." He added that they're continuously benchmarking it and asking it to improve its own performance across open and proprietary models.

## Why the take is about speed

If a small model really goes from nothing to usable in about 3 days of pretraining, iteration cycles shrink fast. @yacineMTB's point lands there: a lab with frontier-scale compute could sweep hyperparameters on a recipe like this over a month. That's the implicit challenge to the bigger labs, and it's why the infra comments are getting more attention than the model itself.

The open question is the weights. They're promised next week, after safety testing. Until they land, the fine-tuning crowd can only use it through Tinker.

## Questions this post answers

### How long did it take to pretrain Thinking Machines' v2 small model?

Pretraining took just under 3 days, or 68 hours, running on GB300 GPUs, according to Thinking Machines engineers. Post-training followed with 17 million rollouts spread over 9 days. The entire pipeline, pretraining, mid-training and post-training, used the NVFP4 numeric format rather than higher precision formats.

_Engineers weighing training infrastructure choices can follow efficiency benchmarks like this one on daily.dev._

### What does it cost to use Thinking Machines' v2 small model on Tinker?

Tinker prices the v2 small model at $0.30 per million input tokens, $1.00 per million output tokens, and $1.50 per million tokens for training. It's positioned as a cheap base model for fine-tuning and customization, with open weights promised for release the following week pending safety testing.

_Teams comparing model pricing before committing to a fine-tuning provider can track updates like this on daily.dev._

## Community take

How the wider developer community reacted, aggregated from 5 discussions and 78 comments across x (as of 2026-11-06).

**TL;DR:** Reaction is largely congratulatory and impressed by the 68-hour pretrain and NVFP4-native pipeline, with technical commenters probing precision choices, infra changes, and fine-tuning practicalities rather than pushing back on the result.

**Sentiment:** 62% positive · 33% mixed · 5% skeptical

**The case for**

- Multiple people express genuine awe at pretraining a ~600B model in under 3 days.
- Running RL rollouts in NVFP4 too is seen as technically interesting since trainer and rollout numerics match, potentially avoiding train/inference mismatch.
- Open-weight availability on Tinker now (with full weights next week) is welcomed, with people already testing fine-tuning and RLVR via the API.

**The pushback**

- Some ask whether the NVFP4-native approach traded away quality versus a BF16 baseline, suggesting lingering uncertainty about precision tradeoffs.
- A past version (v1) is noted as hard to fine-tune and prone to collapsing quickly, raising concern about whether v2 fixes this.
- Questions about what was dropped (e.g., audio support) imply some capability regression versus prior generation.

**By community**

- x (positive): Congratulations and genuine technical curiosity dominate, with only light skepticism about precision tradeoffs and fine-tuning stability.

**Open questions**

- Will NVFP4 training trade off any quality compared to BF16/FP8 baselines, and what exactly is the NVFP4 coverage (QKV, last layers, routed experts)?
- What specifically changed in infra/data pipeline to cut pretraining time to 3 days?
- Why was audio support dropped in this generation?
- Will recommended fine-tuning settings (LR, LoRA rank, batch size) be published given v1's tendency to collapse during fine-tuning?

**Highlights**

> @soumithchintala Very cool that you pretrained in NVFP4. It works well with the right adjustments yet many will default to FP8/BF16 when so much compute is on the line. What would be interesting to know is the NVFP4 coverage (what about QKV? last few layers? or just routed experts?), whether you
> — [counathe on x · 2 points, 1 comments](https://x.com/counathe/status/2118534947824963748)

> @soumithchintala Training natively in NVFP4 raises a question for the weight release: what happens when people requantize? Swapping Gemma 4 between its QAT and PTQ builds, only 23-32% of greedy outputs stayed word-for-word identical, while repeats of one build matched 100/100. Evals per format?
> — [TaikoToeda on x · 1 comments](https://x.com/TaikoToeda/status/2118521555030536682)

> @soumithchintala Doing the RL in NVFP4 too is the interesting part, since rollouts and the trainer end up on the same numerics. Did that help with the usual train/inference mismatch in RL? And will the open weights ship in NVFP4 natively, so 4-bit inference has basically no quant loss?
> — [SoumilBhandari on x](https://x.com/SoumilBhandari/status/2118503112768164215)

> @soumithchintala Soumith, do you guys have recommended settings for LR, Lora rank, batch size etc? V1 was really hard to finetune, the model collapsed pretty quickly…
> — [migtissera on x](https://x.com/migtissera/status/2118507164516171883)

> @soumithchintala Did you run a BF16 baseline at smaller scale to see what NVFP4 cost you in loss? 68 hours for 8.4T tokens is quick, so I'm curious if anything got traded away.
> — [AIQuanting on x · 1 comments](https://x.com/AIQuanting/status/2118492827361636613)

**Source threads**

- [x](https://x.com/soumithchintala/status/2118490925634556026) · 0 points · 50 comments
- [x](https://x.com/cHHillee/status/2118491115896877293) · 0 points · 2 comments
- [x](https://x.com/cHHillee/status/2118492023573577794) · 0 points · 26 comments
- [x](https://x.com/ClementDelangue/status/2118496671520686256) · 0 points · 0 comments
- [x](https://x.com/soumithchintala/status/2118498314371112986) · 0 points · 0 comments

## Community discussion

Top comments from developers on daily.dev.

**@kinurimahoshadhiedirisinghe** · 0 upvotes

> I am like to study ML. So where Can I start?

## Similar posts on daily.dev

- [Fine-Tuning Qwen3-8B with Tinker: From 50% to 76% Accuracy, and Burning Through My Budget](https://daily.dev/posts/fine-tuning-qwen3-8b-with-tinker-from-50-to-76-accuracy-and-burning-through-my-budget-v3s9r5qyt) · DEV · 14 upvotes · 0 comments

---

Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#open-source](https://daily.dev/tags/open-source)

[View this post on daily.dev](https://daily.dev/posts/thinking-machines-v2-small-model-trained-in-68-hours-nvr2rrg0e)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"Thinking Machines' v2 small model trained in 68 hours","url":"https://daily.dev/posts/thinking-machines-v2-small-model-trained-in-68-hours-nvr2rrg0e","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/thinking-machines-v2-small-model-trained-in-68-hours-nvr2rrg0e"},"datePublished":"2026-11-05T23:54:26.860Z","dateModified":"2026-11-06T02:56:16.516Z","description":"Thinking Machines trained its v2 small model in 68 hours on GB300s, using NVFP4 across pre-, mid- and post-training, according to @soumithchintala....","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/thinking-machines-v2-small-model-trained-in-68-hours-nvr2rrg0e","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"machine-learning,open-source","timeRequired":"PT2M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Thinking Machines' v2 small model trained in 68 hours"}]}
{"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/thinking-machines-v2-small-model-trained-in-68-hours-nvr2rrg0e","comment":[{"@type":"Comment","text":"I am like to study ML. So where Can I start?","datePublished":"2026-11-06T00:23:41.256Z","url":"https://daily.dev/posts/nVR2rRg0E#c-uoHDuLqiw","author":{"@type":"Person","name":"Kinuri Mahoshadhi Edirisinghe","url":"https://daily.dev/kinurimahoshadhiedirisinghe","image":"https://media.daily.dev/image/upload/s--eStdL6Z4--/f_auto/v1791302998/avatars/avatar_GO6oJErCIxHbAdVdAwpIF?_a=BAMAMicg0"}}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/thinking-machines-v2-small-model-trained-in-68-hours-nvr2rrg0e#faq","mainEntity":[{"@type":"Question","name":"How long did it take to pretrain Thinking Machines' v2 small model?","acceptedAnswer":{"@type":"Answer","text":"Pretraining took just under 3 days, or 68 hours, running on GB300 GPUs, according to Thinking Machines engineers. Post-training followed with 17 million rollouts spread over 9 days. The entire pipeline, pretraining, mid-training and post-training, used the NVFP4 numeric format rather than higher precision formats. Engineers weighing training infrastructure choices can follow efficiency benchmarks like this one on daily.dev."}},{"@type":"Question","name":"What does it cost to use Thinking Machines' v2 small model on Tinker?","acceptedAnswer":{"@type":"Answer","text":"Tinker prices the v2 small model at $0.30 per million input tokens, $1.00 per million output tokens, and $1.50 per million tokens for training. It's positioned as a cheap base model for fine-tuning and customization, with open weights promised for release the following week pending safety testing. Teams comparing model pricing before committing to a fine-tuning provider can track updates like this on daily.dev."}}]}
```

