@MaximeRivest: Qwen 4B, trained on 500 scientific publication rewrites (so far) for accessibility (see for example), is beating Opus, …
Summary
Maxime Rivest reports that a Qwen 4B model fine-tuned on ~500 scientific publication rewrites for accessibility now outperforms frontier models like Opus, Astra, and Luna — achieved in just 3 hours on a consumer RTX 3090 using coding agents to generate the training set.
View Cached Full Text
Cached at: 10/02/26, 08:38 AM
Qwen 4B, trained on 500 scientific publication rewrites (so far) for accessibility (see 🧵for example), is beating Opus, Astra and Luna (before they were prompt-optimized) and is well on its way to fully overtaking Luna.
This only took 3 hours of running on my 5-year-old 3090 at home.
Fine-tuning is no longer difficult with the help of coding agents and frontier models to create the training set. Try telling your coding agent to do it!
Similar Articles
@MaximeRivest: let's make science more accessible! its free and opensource! https://sciencemadereadable.com
Maxime Rivest launched Science, Made Readable, a free open-source site that rewrites open-access scientific papers into plain language for general audiences, using small fine-tuned Qwen 3.5 models (9B, 4B, 0.8B) that run on a home RTX 3090 and a rented H100. A judge model scored the 9B model at 7.8/10, slightly above GPT-6 Luna (7.5) but below Claude Opus (8.5), at about half a cent per paper.
@RoundtableSpace: Qwen 3.8 27B running locally on an RTX 5090 beats Opus 4.8 in personal benchmarks at up to 200 tokens per second with n…
Qwen 3.8 27B, a 27-billion parameter AI model, outperforms Opus 4.8 in personal benchmarks when running locally on an RTX 5090 GPU, achieving up to 200 tokens per second without internet or API access.
@NFT_Chen: Mind-Blowing! Enhanced Qwen3.8-27B Rivals Opus 5.5, GPT-6 Astra! It's just 27B parameters, yet it undergoes specialized…
A community fine-tune of Qwen3.8-27B (Coder390-EfficientThink) is released, post-trained with SFT + RLOO + MTP on 3.9M lines of code and distilled from frontier models, boosting LiveCodeBench to 90 while cutting reasoning tokens by ~44-53%. Apache-2.0 GGUF builds run locally on 24GB GPUs via llama.cpp or SGLang with DFlash2 speculative decoding.
@itsPaulAi: I can't believe it Qwen3.8-27B is matching Opus 4.6 Max... the model that was the best (and the most expensive) just 6 …
Qwen3.8-27B, a 27-billion parameter AI model with open weights, matches the performance of Opus 4.6 Max and can be run locally on a laptop.
@TheAhmadOsman: Qwen 3.8 27B weights and benchmarks are now out This thing in full precision would fit on a single RTX PRO 6000 and bea…
Qwen 3.8 27B weights and benchmarks are released, reportedly beating Opus 4.6 Max in several benchmarks while fitting on a single RTX PRO 6000 in full precision.