You may have heard that GLM-5.2 at 328 token/s is cool,
How about 392?
Databricks is now #1 in inference speed for GLM-5.2 on Artificial Analysis. It's a great model, and we did a lot of optimizations.
8:26 AM · Jun 25, 2026 · 60K Views
Databricks' Yuchen Jin says system-level optimizations pushed GLM-5.2 inference to a leaderboard-topping 392 tokens/s
Story Overview
Databricks engineer Yuchen Jin highlighted system-level tweaks that lifted GLM-5.2 inference to 392 tokens per second on the Artificial Analysis leaderboard, surpassing Fireworks' prior 328 tokens per second mark on a custom endpoint. The result spotlights engineering refinements applied to a leading open-weights model that already posts strong intelligence scores and a 1M-token context window.
78676639381.4K
Developer Impact
Serving speed directly shapes real-world usability
Faster output rates can cut latency and raise throughput for developers running GLM-5.2, yet the announcement supplies no workload-specific benchmarks or cost figures to quantify the advantage.
Open Question
The path to these gains remains under the hood
The claim credits extensive optimizations and references the OpenCode harness, but leaves hardware details, batch settings, and exact methods unspecified, so independent replication stays an open question.
Sentiment
Many users are excited about Databricks topping the GLM-5.2 inference speed leaderboard at 392 tokens/s because it highlights impressive open-source progress and future AI potential.
Pos
94.1%
Neg
5.9%
18 comments with sentiment.
Cluster Engagement
Digg Deeper
No Digg Deeper questions have been answered for this story yet.
Posts from X
Most Activity
Most Activity
VIEWS8.5KLIKES89RETWEETS9
Matei Zaharia@matei_zaharia
The Databricks inference team is 🔥
Yuchen Jin@Yuchenj_UW
You may have heard that GLM-5.2 at 328 token/s is cool,
How about 392?
Databricks is now #1 in inference speed for GLM-5.2 on Artificial Analysis. It's a great model, and we did a lot of optimizations.
2hViews 8.5KLikes 89Bookmarks 9
BOOKMARKS10REPLIES4
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Does this answer your question about its positioning @TheZvi ? I told you it's a hardware issue. GLM 5.2 can be served very, very quickly. Databricks won't stop here, neither will their competitors. Chinese models are *much better* than Chinese silicon.
Yuchen Jin@Yuchenj_UW
You may have heard that GLM-5.2 at 328 token/s is cool,
How about 392?
Databricks is now #1 in inference speed for GLM-5.2 on Artificial Analysis. It's a great model, and we did a lot of optimizations.
2hViews 7.2KLikes 31Bookmarks 10
Yuchen Jin@Yuchenj_UW
Databricks is all in oss models.
We're looking for cracked AI inference engineers to join us at Databricks AI to produce trillions of tokens.
DM me if you have: - Contributed to open-source ML systems like SGLang/vLLM/PyTorch/Dynamo - Experience serving LLMs at large scale
Yuchen Jin@Yuchenj_UW
You may have heard that GLM-5.2 at 328 token/s is cool,
How about 392?
Databricks is now #1 in inference speed for GLM-5.2 on Artificial Analysis. It's a great model, and we did a lot of optimizations.
2hViews 2.8KLikes 53Bookmarks 9
Yuchen Jin@Yuchenj_UW
@teortaxesTex @TheZvi chip matters, but so do the optimizations we made across speculative decoding, kernels, etc. Proud of the Databricks inference team.
We had reached 201 tok/s on H200s a few days ago, and B300s gave us a substantial additional boost.
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Does this answer your question about its positioning @TheZvi ? I told you it's a hardware issue. GLM 5.2 can be served very, very quickly. Databricks won't stop here, neither will their competitors. Chinese models are *much better* than Chinese silicon.
2hViews 532Likes 14Bookmarks 2
kache@yacineMTB
Zephyr@zephyr_z9
@teortaxesTex @TheZvi So Databricks hired Yuchen to build their inference provider biz
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Does this answer your question about its positioning @TheZvi ? I told you it's a hardware issue. GLM 5.2 can be served very, very quickly. Databricks won't stop here, neither will their competitors. Chinese models are *much better* than Chinese silicon.
1hViews 2KLikes 4Bookmarks 1
kache@yacineMTB
@Yuchenj_UW awesome
Yuchen Jin@Yuchenj_UW
You may have heard that GLM-5.2 at 328 token/s is cool,
How about 392?
Databricks is now #1 in inference speed for GLM-5.2 on Artificial Analysis. It's a great model, and we did a lot of optimizations.
2hViews 1.5KLikes 11Bookmarks 0
Ai agent@ai_agent001
@Yuchenj_UW 392 now, calling it 400+ by next week
2hViews 266Likes 5
Yuchen Jin@Yuchenj_UW
@ai_agent001 it's only a start.
2hViews 286Likes 7
Taelin@VictorTaelin
@Yuchenj_UW Quantized?
Yuchen Jin@Yuchenj_UW
You may have heard that GLM-5.2 at 328 token/s is cool,
How about 392?
Databricks is now #1 in inference speed for GLM-5.2 on Artificial Analysis. It's a great model, and we did a lot of optimizations.
1hViews 1.1KLikes 5Bookmarks 0
Yuchen Jin@Yuchenj_UW
@yacineMTB wow, great to see our Unity AI Gateway page on X.
We’ve launched to a select people/workspaces and plan to roll it out more broadly.
btw, how do you feel about our AI gateway UI/UX?
2hViews 105Likes 3
Yuchen Jin@Yuchenj_UW
@matei_zaharia LFG!
Matei Zaharia@matei_zaharia
The Databricks inference team is 🔥
1hViews 293Likes 6Bookmarks 0
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
@Yuchenj_UW @TheZvi Yeah, I wonder how much faster zAI could go with fixed hardware. But you have structurally higher ceiling thanks to Blackwell.
Yuchen Jin@Yuchenj_UW
@teortaxesTex @TheZvi chip matters, but so do the optimizations we made across speculative decoding, kernels, etc. Proud of the Databricks inference team.
We had reached 201 tok/s on H200s a few days ago, and B300s gave us a substantial additional boost.
2hViews 244Likes 6Bookmarks 0
goodalexander@goodalexander
@Yuchenj_UW holy shiet
1hViews 319Likes 3
Zvi Mowshowitz@TheZvi
@teortaxesTex Certainly helps since they're charging the same price.
1hViews 163Likes 2
Aishwarya Goel (Ash)@aishwarya_08
@Yuchenj_UW Things are getting out of hand now 😂
2hViews 135Likes 2
kache@yacineMTB
@Yuchenj_UW get started with databricks is unecessary. remove that hop. it will increase your conversion by 30%
2hViews 87Likes 2
Kai Benetti@kai_benetti
@Yuchenj_UW Funny how fast this flips, whoever's on top here barely holds it for a day
2hViews 221Likes 1
goodalexander@goodalexander
@Yuchenj_UW where do i buy
1hViews 220Likes 1
Manishhh@iamanishx
@Yuchenj_UW B300s and speculative decoding go brr
2hViews 96Likes 1
Load more posts