Unverified50% confidenceBenchmarkExact time
After integrating local Qwen3 into the DeepSeek Harness framework, thinking mode runs at approximately 30 tokens per second, while disabling thinking mode peaks at close to 40 tokens/s for single tasks.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/5/2026
First Seen
Valid until: 1/3/2027
Sources
Related Entities
Related Claims
Partially VerifiedBoth Qwen 3 Max and DeepSeek-V3 adopt the MoE architecture with total parameter counts reaching hundreds of billions, but the parameters actually activated during inference are only a small fraction of the total.67% similarUnverifiedQwen3.7 Max already supports context lengths at the million-token level.66% similarUnverifiedA user with a private benchmark stated that Qwen 3.8 was only the second local model able to correctly reason through their test questions, but it took 12 minutes and 30 seconds and used five times as many tokens as Gemma 4.65% similarUnverifiedQwen3 27B natively supports a context of 262,000 tokens and can be extended to 1,000,000 tokens via YaRN extrapolation.64% similarUnverifiedQwen3 27B achieved a 100% pass rate in the 256K long-context memory retrieval test (needle in a haystack), with all 15 runs being accurate.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/978648API
curl https://kongchang.com/api/v1/knowledge/claims/978648MCP
get_claim(id=978648)