explainx / blog
VibeThinker 3B, based on Qwen 2.5-Coder, reportedly achieves Claude Opus 4.5-level performance at 3 billion parameters — through post-training refinements including distillation from RL checkpoints and a final RL-based instruct pass. Here is what the paper says and why it matters for local AI.
Jul 3, 2026
GitHub Copilot now offers Moonshot AI's Kimi K2.7 Code as a selectable open-weight model — the first in Copilot's model picker. Pro plans first; Business and Enterprise require admin enablement. Here's how to turn it on.
Jun 29, 2026
Cline bundles GLM-5.2 and Chinese open-weight APIs for $9.99/mo — no key juggling. A $1.99 intro promo runs through npm install. Quota is opaque; here's what we know.
Jun 29, 2026
After HN front-page hype, hands-on tests say Qwen 3.6 27B dense is the local sweet spot — better code than the 35B MoE, runnable at Q8 on 48GB RAM. Full llama.cpp + OpenCode config inside.
On June 16, 2026, AI researcher Ahmad (@TheAhmadOsman) shared a paper that stopped a lot of people mid-scroll: VibeThinker 3B, a 3-billion-parameter model based on Qwen 2.5-Coder, reportedly achieves performance comparable to Claude Opus 4.5 on coding benchmarks.
Ahmad's January 2026 prediction, now resurfacing:
"We will have Claude Code + Opus 4.5 quality (not nerfed) models running locally at home on a single RTX PRO 6000 before the end of the year."
VibeThinker 3B, if the claims hold, is that model — six months ahead of schedule.
The paper is available at arXiv:2606.16140 and has been gaining traction in the local AI community.
Since the original paper dropped, the community has been running evaluations. The numbers now circulating widely:
| Benchmark | Score |
|---|---|
| AIME 2026 | 94.3 |
| LiveCodeBench v6 (Pass@1) | 80.2 |
| LeetCode contests (unseen) | 96.1% |
Chubby (@kimmonismus), whose AI newsletter reaches 225K+ subscribers, flagged these as "crazy" — and the reaction on X broadly agreed. A 3B model at 96.1% on unseen LeetCode contests is the kind of result that makes you check whether the benchmarks were in the training set.
The community verdict from the thread: these results appear legitimate but domain-specific. "Certain forms of verifiable reasoning may be highly compressible into small dense models. Frontier-scale models still matter for broad knowledge and general-purpose capability, but compact reasoning models are becoming a serious complementary path."
VibeThinker 3B is not a new pretraining run. It starts from Qwen 2.5-Coder — Alibaba's open-source code-focused model family — and applies a four-stage post-training pipeline (now clearer from community analysis):
Stage 1: Curriculum SFT Supervised fine-tuning on a difficulty-ordered dataset — easy problems first, progressively harder. This builds baseline instruction-following before RL begins.
Stage 2: Multi-domain RL Reinforcement learning across multiple coding domains simultaneously (not just Python function completion) — the model gets execution feedback from tests across different problem types.
Stage 3: Offline self-distillation The model generates its own high-quality solutions, filters them by correctness, and trains on them — a form of self-improvement without requiring a larger teacher.
Stage 4: RL-based instruct tuning A final RL pass focused on instruction-following — improving the model's ability to handle complex prompts and produce structured outputs.
As researcher Francesco Bertolotti (arxiv:2606.16140) noted when the paper first landed: "These results were achieved primarily through post-training refinements on Qwen2.5-Coder. The paper doesn't provide many details, but it appears they distill from RL ckpts and then do a final RL-based instruct RL." The fuller four-stage picture above has since been pieced together from the paper and community analysis.
The reflexive skepticism from the community — "if 3B matches Opus 4.5, how come 200B+ open-source models aren't at that level?" — is worth taking seriously.
The answer is domain specificity and post-training. There are several things to understand:
1. Benchmark ≠ general capability Opus 4.5 is a general-purpose frontier model. If VibeThinker 3B matches it on specific coding benchmarks, that does not mean it matches across all domains. A 3B model that is extraordinarily good at Python function completion may outscore Opus 4.5 on HumanEval while being far behind on anything that requires long-context reasoning, general knowledge, or multi-domain synthesis.
2. RL post-training is disproportionately effective for specific tasks Reinforcement learning from code execution feedback — running code, checking whether it passes tests, using pass/fail as a reward signal — is remarkably sample-efficient for improving coding performance. A small model with well-targeted RL post-training can punch significantly above its parameter weight on the specific tasks the RL was optimised for.
3. 200B+ open-source models are typically not RL-fine-tuned this aggressively for single tasks Large open-source models are typically trained for breadth. A 3B model optimised exclusively for coding through RL may genuinely outperform a 200B general-purpose model on coding-specific benchmarks.
4. Distillation from RL checkpoints is a relatively new technique The specific combination — distill from a model that has already done RL, then do your own RL pass — is not widely tested at scale. VibeThinker 3B may be one of the earlier papers to show this approach working this well.
At 3 billion parameters, VibeThinker 3B sits comfortably in the range of locally deployable models:
An RTX 3090, RTX 4090, or RTX PRO 6000 can all run this model at full precision. A mid-range consumer GPU (RTX 4060 Ti) can run it quantised.
If the benchmark numbers translate to real-world coding quality, this means developers who want Claude Code-quality coding assistance can run it on local hardware — no API costs, no rate limits, no data leaving the machine. That is a meaningful change in the economics and privacy calculus of AI-assisted development.
The community response on X has been appropriately divided. Nicolás Canala asked the question directly: "benchmaxxed or actually getting good outputs?"
This is the right question. Models can achieve high benchmark scores through:
Without independent evaluation on tasks that weren't in the training distribution, it's impossible to know whether VibeThinker 3B's performance is generalisable or benchmark-specific. Francesco Bertolotti's note that the paper "doesn't provide many details" about the methodology is a flag worth taking seriously.
The claim deserves independent replication. The community is watching.
VibeThinker 3B is part of an accelerating trend: the gap between frontier models and small locally-runnable models is closing faster than most predictions anticipated.
The pattern:
If this trajectory continues, the argument for always hitting frontier model APIs for coding tasks weakens significantly. Local models with narrow task specialisation — especially when combined with agent loop architectures that use fast local inference for verification passes — become a genuinely competitive alternative.
This is also relevant context for the SpaceX acquisition of Cursor — the value of an AI coding tool is partly in model access, but also in workflow, context management, and IDE integration. If the models commoditise, the workflow layer becomes more important.
