Quasar 438B: Europe's Leading AI Model
TL;DR Highlight
Quasar 438B claims Europe's top AA index score (43) with fast responses
Who Should Read
A backend developer comparing European or API-only reasoning models for agent and coding workloads
Core Mechanics
- Top European score of 43 on AA index
- It scores 43 on the Artificial Analysis Intelligence Index v4.1.1, a composite of nine evaluations.
- Mistral Medium 3.5 scores 30, Nemotron 3 Ultra 38, and Inkling 42.
- Claude Opus 5 leads at 63.
- 500 tokens in 15.3 seconds
- This includes thinking time.
- Only three models in the comparison are faster, and only Gemini 3.7 Flash (11.5s, 56) also scores higher.
- Mistral Medium 3.5 takes 18.8s and Inkling 48.3s.
- Long-context reasoning near the frontier
- It scores 75.0 on AA-LCR, within a point of Claude Opus 5 (75.7).
- It beats Nemotron 3 Ultra by 4.0 points and Mistral Medium 3.5 by 9.7.
- Terminal agent work still has a big gap
- It scores 69.3 on Terminal-Bench v2.1.
- That is 18.7 points above Mistral Medium 3.5 and 15.4 above Nemotron 3 Ultra.
Community Reactions
- Closed weights make benchmarks hard to trust, some said
- One commenter asked whether the weights are public.
- Another said an API-only model will get little uptake and be hard to verify.
- Opinion that it is a modified GLM 5.2 or MiniMax M3
- Commenters noted the company's core business is removing parameters from models.
- They pointed to the changelog saying capabilities match GLM 5.2, the same effort settings (high/max), and text-only support.
- The commenter admitted they might be wrong.
- Point that a 438B model trails Qwen3.8-27B
- Some asked whether the benchmarks were gamed.
- Others called it ironic for a larger model to lose to a smaller one.
- A harness added about 10% on SWE Pro, one commenter reported
- They put a proxy in front of Qwen3.8-27B that nudged it to continue or retry on bad output.
- It also detected repetition and reminded the model about tool use.
- Their point: open models are tested naked, while frontier models are tested with harnesses.
How to Apply
- Build an index-vs-latency table for your candidates
- In 30 minutes, pull each candidate's index score and 500-token response time from Artificial Analysis.
- Compare against your service's latency budget and filter the list.
- Evaluate on your own data
- For closed-weight models, run 20-50 of your own test cases and measure accuracy and latency.
- Use the free first month of the CompactifAI API to do it.
- Add a retry/reminder harness to your agent
- Detect empty output, malformed JSON, and repeated tool calls, then inject retry messages.
- Compare scores with and without the harness before switching models.
Try It
# Paste into Claude or ChatGPT. A simplified version of the harness idea from a comment (not the article's method)
You are an agent that uses tools. Follow these rules.
- If you get stuck, do not stop; try a different approach.
- If your output is empty or JSON is malformed, detect it yourself and answer again better.
- If you call the same tool repeatedly, re-plan first.
- If you haven't used tools for a while, remember that you can.
Task: {{task description}}
Next step: run the same task once more without these rules and compare the results.Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.