Is Quark state of the art?Pending benchmarks.Announced Oct 23, 2026; no independent results yet.
FirstPrinciples' 9B science and maths reasoning fine-tune of Qwen3.5-9B (SFT on a curated physics, maths and coding mix). Open weights on Hugging Face under Apache 2.0, with a matching Q8_0 GGUF (FirstPrinciplesLabs/quark_quantized_weight_only-Q8_0) and Ollama tags. Also available on FirstPrinciples' Theo research platform. Not a conversation or multimodal chat model: the card says it was trained only on single-turn science, maths and code.
The details so far
- Lab
- FirstPrinciples
- Announced
- Oct 23, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 9B
- Role
- Reasoning model for physics, science and maths; step-by-step solutions with a long thinking budget
- Base
- Fine-tune of Qwen/Qwen3.5-9B (HF base_model); separate from the Qwen3.5 9B catalog page
- Training (lab)
- About 200,000 curated physics, science, maths and coding problems with worked solutions; rejection sampling; n-gram decontamination against named benchmark sets
- Context / output
- Lab recommends up to 131,072 output tokens in thinking mode (Qwen3.5 sampling settings)
- Checkpoints
- FirstPrinciplesLabs/quark (bf16) and FirstPrinciplesLabs/quark_quantized_weight_only-Q8_0 (GGUF) on Hugging Face; Ollama firstprinciples/quark and FirstPrinciples/quark-q8
- License
- Apache 2.0
- Platform
- Also on Theo Platform (FirstPrinciples research environment); local Transformers / vLLM / llama.cpp paths documented on the card
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: GPQA Diamond 87.4 (bf16) and 87.8 (Q8_0) vs 81.6 for Qwen3.5-9B under the lab's EvalScope/vLLM setup (thinking mode).
Lab claim: AIME 2025 91.7 (bf16) and 94.0 (Q8_0) vs 86.9 for the base; HLE text-only 22.7 (bf16) and 22.3 (Q8_0) vs 15.9 for the base.
Lab claim: Q8_0 GGUF matches bf16 on every benchmark in the published table.
Sources
REPOFirstPrinciplesLabs/quark model card (Hugging Face)huggingface.co/FirstPrinciplesLabs/quarkREPOFirstPrinciplesLabs/quark_quantized_weight_only-Q8_0 (Hugging Face)huggingface.co/FirstPrinciplesLabs/quark_quantized_weight_only-Q8_0LABFirstPrinciples on X: Quarkx.com/FirstPrincip1es/status/2118733991398322177 · Nov 6, 2026Tracking continues until an independent board lists it.