VibeThinker 3B
Hacker ApprovedRanked 2 of 4 reasoning models by recommenders
Small open model that people recommend for math and logical reasoning, performing well above its size. Common caveats: weak outside its specialty, poor at structured output and tool use.
Small open model that people recommend for math and logical reasoning, performing well above its size. Common caveats: weak outside its specialty, poor at structured output and tool use.
38 people gave an opinion
- recommend it
- 28
- criticize it
- 10
- approval share
- 74%
Recommended in 4 of 13 months
- November 2025: recommended
- December 2025: not recommended
- January 2026: not recommended
- February 2026: not recommended
- March 2026: not recommended
- April 2026: not recommended
- May 2026: not recommended
- June 2026: recommended
- July 2026: recommended
- August 2026: recommended
- September 2026: not recommended
- October 2026: not recommended
- November 2026: not recommended (month in progress)
Why it carries the badge
- 28 people recommend itNeeds 10 or more, met.
- 2.8 recommenders for every criticNeeds 2 to 1 or better, met.
- Recommended in 4 of 13 monthsNeeds 4 or more, met.
The same three rules apply to every product, and the badge can’t be bought. How the badge works
What people say about it
The reasons people give, with the number of comments that give each one. Open a reason to read what they said.
Why people recommend it
Each mark is one comment
Strong math and logical reasoning
“basically a match for the larger frontier models -- hundreds of times larger -- in the narrow domain of mathematical and logical reasoning problems”
Comment from 17 Aug 2026 in “Qwen3.8 27B scores 52 on Artificial Analysis” “performs like models 100x its size on mathematical and logical reasoning tasks”
Comment from 23 Aug 2026 in “Fable and the end of the free lunch” “It's a densely trained on math whuch let's it punch way higher than it should”
Comment from 20 Jun 2026 in “GPT-5.5 hallucinates 3x more than MIT-licensed GLM-5.2” - Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO”
- Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO”
Strong results for its size
“basically a match for the larger frontier models -- hundreds of times larger -- in the narrow domain of mathematical and logical reasoning problems”
Comment from 17 Aug 2026 in “Qwen3.8 27B scores 52 on Artificial Analysis” “on well-constrained, verifiable reasoning tasks, first-tier performance is no longer the exclusive domain of ultra-large models”
Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO” “Overall it still seems extremely good for its size and I wouldn't expect anything below 30B to behave like that.”
Comment from 13 Nov 2025 in “VibeThinker-1.5B” - Comment from 14 Nov 2025 in “VibeThinker-1.5B”
- Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO”
Writes working Python code
“It can seemingly reliably write working Python code though”
Comment from 17 Jun 2026 in “VibeThinker-3B achieves 80.2 on LCBv6” “It seems weirdly good at some simple python questions.”
Comment from 13 Nov 2025 in “VibeThinker-1.5B” Finds valid solutions to problems
“solving unformalized problems like VibeThinker does”
Comment from 24 Aug 2026 in “Fable and the end of the free lunch” “And surprisingly it found a valid solution!”
Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO”
What people warn about
Each mark is one comment
No tool calling capability
“it's trained to generate long CoT to reason through problems (and does it well!) but has ~no tool calling capability”
Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO” “I'm not seeing any mention of tools in the paper, much less a bias towards "curiosity" to use those tools”
Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO” “This model was not trained on tool-calling or agent-based programming data.”
Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO” Rambling thinking, loops, wrong steps
“beside a marked tendency towards rambling thinking, it's just plain implausibly good for a 1.5B model”
Comment from 14 Nov 2025 in “VibeThinker-1.5B” “Some incorrect things and getting into thinking loops, but always coherent”
Comment from 13 Nov 2025 in “VibeThinker-1.5B” “it takes so many shortcuts it writes wrong steps in its answers”
Comment from 17 Jun 2026 in “VibeThinker-3B achieves 80.2 on LCBv6” Limited use outside math
“it failed horribly just showing me a rectangle and a black circle”
Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO” “the real world use cases for this model seem to be very limited”
Comment from 13 Nov 2025 in “VibeThinker-1.5B” Benchmark gains look like overfitting
“It just hardcoded the tasks for bench, It does even respond normally”
Comment from 23 Jun 2026 in “VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO” “Overfitting, no need to argue about anything I think?”
Comment from 18 Jun 2026 in “Why Weibo's tiny VibeThinker-3B has the AI world arguing over benchmarks again”
Other reasoning models people recommend
The three most recommended other reasoning models, with the same counts for each, so you can compare like for like. The tick on each bar marks the 2 to 1 line the badge needs.
| Product | Recommend | Criticize | Approval | Months | Owned | Badge |
|---|---|---|---|---|---|---|
| OpenAI o3 | 37 | 13 | 74% | 11 | — | Approved |
| VibeThinker 3BThis page | 28 | 10 | 74% | 4 | — | Approved |
| DeepSeek R1 | 25 | 15 | 63% | 10 | — | — |
| OpenAI o1 | 12 | 10 | 55% | 7 | — | — |
What people said
40 of the 53 comments that mention it, newest first. Each links to the original comment.
“it's very weak outside its specialty”
“VibeThinker 3B constitutes extraordinary evidence, IMO. The first such evidence I've seen myself.”
“solving unformalized problems like VibeThinker does”
“It's impressive that it does what it does”
“performs like models 100x its size on mathematical and logical reasoning tasks”
“basically a match for the larger frontier models -- hundreds of times larger -- in the narrow domain of mathematical and logical reasoning problems”
“this is great! thanks for sharing”
“It's a model trained specifically for reasoning. Basically a problem solver.”
“IMO VibeThinker is the most interesting open model since the OG DeepSeek R1”
“I think vibethinker is heavily overtrained on not attempting to solve open problems.”
“this model provides something of a lower bound on just how much world knowledge is really needed for unrelated reasoning tasks”
“It is crazy good at math.”
“Sounds like something that could be pretty useful as a 'validation' subagent.”
“This model was not trained on tool-calling or agent-based programming data.”
“this might actually become my main work horse”
“It just hardcoded the tasks for bench, It does even respond normally”
“it's trained to generate long CoT to reason through problems (and does it well!) but has ~no tool calling capability”
“boy, does it sure reason through math problems well”
“on well-constrained, verifiable reasoning tasks, first-tier performance is no longer the exclusive domain of ultra-large models”
“So it seems there is definite progress here.”
“Seems like a really good model to use in an IDE when you still want control over the code structure”
“this is a good model. I benchmark reasoned answers to qwen 3.6 27b (no think)+ bash and it held up.”
“That plus this model should give you a very powerful and focussed assistant.”
“You can't get 94.3 on AIME26 knowing nothing.”
“And surprisingly it found a valid solution!”
“I'm not seeing any mention of tools in the paper, much less a bias towards "curiosity" to use those tools”
“the size-to-capability ratio is good”
“It's not great on structured output”
“it failed horribly just showing me a rectangle and a black circle”
“It's a densely trained on math whuch let's it punch way higher than it should”
“Overfitting, no need to argue about anything I think?”
“It can seemingly reliably write working Python code though”
“it takes so many shortcuts it writes wrong steps in its answers”
“beside a marked tendency towards rambling thinking, it's just plain implausibly good for a 1.5B model”
“the real world use cases for this model seem to be very limited”
“Overall if you're memory constrained it's probably still worth to try and fiddle around with it”
“Some incorrect things and getting into thinking loops, but always coherent”
“Just doesen't seem to work at a useable level.”
“Overall it still seems extremely good for its size and I wouldn't expect anything below 30B to behave like that.”
“It seems weirdly good at some simple python questions.”
Counts come from 1,795,640 Hacker News comments, November 2025 to November 2026. How we count. Report a mistake.