Unverified50% confidenceBenchmarkExact time
Developer Primitive reported: 91.2 tokens/s with speculative decoding off, reaching 142.6 tokens/s after enabling 3 speculative tokens, approximately a 56% speedup under the same configuration.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/28/2026
First Seen
Valid until: 12/27/2026
Sources
Related Entities
Related Claims
VerifiedUnder a 4×RTX 3090 SXM4 configuration, the quantized version of GLM-5.2 achieves a token generation speed of approximately 6.3–6.8 tokens/second and a prompt evaluation speed of approximately 10.6 tokens/second.77% similarUnverifiedClient-side observations show that the fix task output speed is approximately 242 tokens/second and the Web task speed is approximately 269 tokens/second, with input cache hit rates of approximately 92.5% and 97.8% respectively.73% similarUnverifiedIn prefill speed tests, Halo achieved approximately 650 tokens/s while Spark achieved approximately 2000 tokens/s, with Spark being nearly three times faster.73% similarUnverifiedWhen running Qwen3-Next-Flash (Autoround W4A16 quantization), the prefill speed was approximately 1.3k tokens/s, and the generation speed reached 70 tokens/s (code) / 60 tokens/s (prose), with support for 128k+ long context.71% similarUnverifiedUnder this configuration, the 35B model achieves a measured single-user peak speed of approximately 80 tokens/second, with concurrent throughput exceeding 500 tokens/second.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/956793API
curl https://kongchang.com/api/v1/knowledge/claims/956793MCP
get_claim(id=956793)