title: "3B Matches 600B" — I Was Silent for 5 Minutes After Reading Sina's Paper date: 2026-06-29 platform: toutiao status: finalized cover: cover.png

"3B Matches 600B" — I Was Silent for 5 Minutes After Reading Sina's Paper

Yesterday, Sina (

💡 What You Will Learn

title: "3B Matches 600B" — I Was Silent for 5 Minutes After Reading Sina's Paper date: 2026-06-29 platform: toutiao status: finalized cover: cover.png "3B Matches 600B" — I Was Silent for 5 Minutes A

📜 Table of Contents

• 1. What Exactly Is VibeThinker-3B
• 2. The "Truth" Behind 4 Benchmarks — 3 Up, 1 Down
• 3. Plain-English Summary: What Does This Mean for You?
• 4. My Take
• 5. Let's Talk in the Comments

title: "3B Matches 600B" — After Reading SINA's Paper, I Was Silent for 5 Minutes date: 2026-06-29 platform: toutiao status: finalized cover: cover.png

"3B Matches 600B" — After Reading SINA's Paper, I Was Silent for 5 Minutes

Yesterday, SINA (yes, the portal site SINA) open-sourced a small model with 3B parameters, called VibeThinker-3B.

Let me throw a few numbers at you first — on AIME26, a benchmark at math-olympiad difficulty, it approached DeepSeek V3.2, Qwen3-235B, and other large models with 200 to 333 times its parameter count; on LeetCode contest problems, it solved 123/128, leaving GPT-5.2 and Kimi K2.5 in the dust.

Note: I said "approached," not "matched."

Clickbait headlines love these two words — "3B crushes 600B" — but when you lay the data out flat, that's not what's happening at all.

1. What Exactly Is VibeThinker-3B

It wasn't trained from scratch by SINA. Its base is Alibaba's Qwen2.5-Coder-3B, a 3B-parameter foundation model fine-tuned specifically for code and reasoning.

What SINA did was apply a "three-stage post-training" pipeline: - SFT (Supervised Fine-Tuning): First, teach it to "think step by step" using high-quality math/code problems - Reinforcement Learning: Let the model "score points" through self-play, rewarding correct answers - Self-Distillation: Distill the model's own multi-solution problem outputs into training data

These three steps aren't original to SINA — DeepSeek R1 and Qwen3 have all done this. But SINA's bet is: post-training only, base model untouched, and a 3B can still reach the top.

2. The "Truth" Behind 4 Benchmarks — 3 Up, 1 Down

Straight to the data (source: The Decoder's report + SINA's official card):

See that? "3 up, 1 down" — it can go toe-to-toe with 600B models on logical reasoning tasks, but it's still illiterate on knowledge-intensive tasks.

That's the core finding of SINA's paper, and they gave it a name: the "Parameter Compression-Coverage Hypothesis": - Logical reasoning relies on a few compressible patterns (a few reasoning paths that, no matter how much you distill, stay the same) → a 3B can fit them - World knowledge is encyclopedic coverage ("What year did WWII end?" "What's the 51st element on the periodic table?") → a 3B can't fit it; it needs large parameters

What this means: A 3B model can catch up to large models on the "thinking" level, but it will always be an elementary schooler on the "knowing" level. Don't get excited by "3B crushes 600B" and rush to buy small models — ask it to explain "Trump's second-term cabinet reshuffle," and it will still hallucinate.

Let me add one more detail to make the "3 up, 1 down" table more concrete —

• What AIME26 is: An extended problem set from the American Invitational Mathematics Examination, used by the AI community as a hard yardstick for "olympiad-level reasoning." VibeThinker-3B matching 600B-class models here shows that the logic-chain compression path really works
• What GPQA-Diamond is: A PhD-level science Q&A benchmark covering physics/chemistry/biology and other high-barrier knowledge. The 3B lags far behind because these questions require "knowing" before "reasoning" — which exactly confirms the "knowledge can't be compressed" hypothesis
• What LeetCode 123/128 is: Real-world algorithm contest problems, the "can it actually be used" metric developers care about most — a 3B beating GPT-5.2 here is a real signal

3. Plain-English Summary: What Does This Mean for You?

If you're a developer: - Fits locally: A 3B model runs at 30+ tokens/sec on consumer GPUs (like a 4090) — it can even run on a phone - Right use cases: Code assistance, math problem assistance, algorithm problem assistance → pick VibeThinker-3B - Don't use it for: Writing news, writing summaries, tasks that require "knowing" a lot of facts → pick a large model

If you're a regular user: - You won't use it this month: Open source is for developers; it's far from the Doubao or Kimi in your phone - But your phone's future is being rewritten by it: A 3B that can reason locally means future phones, smart homes, and cars can all run "thinking AI" without uploading every sentence to the cloud — this direction matters a hundred times more than VibeThinker-3B itself

If you're an AI industry observer: - The small-model path has officially moved from "exploration" to "maturity" — DeepSeek R1's distilled small models, Qwen3's small sizes, Llama 3.2 1B/3B, Phi-4 mini, and now SINA's VibeThinker — every major player is betting on "small but sharp" - The large-model path won't die, but general large models + specialized small models will become the mainstream combo

4. My Take

I have a contrarian view —

"3B crushes 600B" is marketing speak. What SINA is really trying to say isn't "small model miracle" at all.

What SINA is saying: In the next 5 years, the AI running in your phone, your car, and your appliances won't be a simplified version of a cloud-based 600B model — it'll be a local 3B small model specifically optimized for "thinking."

Behind this is a battle for compute sovereignty:

• Right now, all AI runs in the cloud → NVIDIA rakes it in, and ordinary people pay every time they ask AI a question
• In the future, 3B models go local → compute sinks to the edge, AI inference costs hit zero, and new scenarios explode

VibeThinker-3B is a milestone event on this path, not a "small model miracle."

And here's the thing — VibeThinker-3B is open source, under the MIT license. That means any phone maker, appliance maker, or car maker can use it directly without paying licensing fees. This is the second time a Chinese company has set a global example of "model democratization," after DeepSeek V3/R1.

5. Let's Talk in the Comments

Here's a question:

Would you rather have a local 3B "mini AI" running on your phone (fast, offline, privacy-safe, but with less knowledge than the cloud), or keep using cloud-based large models (full knowledge, but requires internet, costs money, and may leak privacy)?

Drop your thoughts in the comments — do you care more about "speed + privacy" or "completeness + intelligence"?