Expired50% confidenceClaimExact time
A Reddit technical post claims that a model with only 0.8B parameters achieved a score of 42.15% on the ARC-Challenge benchmark, surpassing the untuned stock Qwen3.5-2B base model and outperforming a solution using SFT distillation with 25,000 data entries.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/6/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 10/20/2026(expired)
Sources
Related Entities
Related Claims
UnverifiedResearch shows that 30B-parameter models have a success rate 15–30 percentage points lower than GPT-4/Claude-3.5 on complex reasoning tasks.68% similarUnverified研究表明当模型参数规模达到3B以上时，1.58-bit模型在多项基准测试中表现可接近甚至匹配同等规模的全精度模型66% similarUnverifiedActual testing of two quantization models: IQ2_XXS achieves 48 tokens/s for short Q&A and 63 tokens/s for long text; IQ3_XXS achieves 30 tokens/s for short Q&A and 39 tokens/s for long text, with VRAM usage of approximately 16.44 GB, activating 10,138 experts—IQ2 is about 60% faster than IQ3.64% similarUnverifiedOn BU Bench V1, the 4B-parameter SFT model surpassed the 45% score reported by DeepSeek V4 Pro and Kimi K2.6 under the open-source Browser Use framework.64% similarUnverifiedA 70B-parameter model requires approximately 35–40GB of VRAM after Q4 quantization, which a single T4 cannot handle, but dual T4s can basically satisfy.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/980499API
curl https://kongchang.com/api/v1/knowledge/claims/980499MCP
get_claim(id=980499)