Wafer · fp4

Serving GLM 5.3. Checked Nov 11.

On watch

Same answer as the maker

typical for the others

56%

57 of 102 prompts

72%

typical for the others

Z.AI · fp8makersame answer twice 62% of the time

Novita · fp875%76 of 102 prompts · typical 71%

Makora · fp474%56 of 76 prompts · typical 71%

Decart · fp474%75 of 102 prompts · typical 71%

GMICloud · fp872%73 of 101 prompts · typical 71%

Reka71%72 of 101 prompts · typical 72%

SiliconFlow · fp870%71 of 102 prompts · typical 72%

Wafer · us58%59 of 101 prompts · typical 72%

Wafer · fp456%57 of 102 prompts · typical 72%

Tenstorrent · fp845%9 of 20 prompts · typical 60%

Every provider gets the same short prompts at temperature 0, so settings cannot change the answer. Honest providers still differ a little, which is why we compare with what is typical, not with 100%.

Question by question

Z.AI · fp8 (maker)Novita · fp8 Makora · fp4 Decart · fp4 GMICloud · fp8 Reka SiliconFlow · fp8 Wafer · us Wafer · fp4

key75%74%74%72%71%70%58%56%

106 questions. Each square is one question. Lit means the same answer as Z.ai. Hover or tap a column.

History

one cell per daily check

Listing

declared precision: fp4
price per 1M: $0.179 in · $1.4 out

Check	How	Measured	Expected	Result
Top answer	106 prompts at temperature 0. How often the top answer equals Z.AI's own endpoint, next to how often the other providers do	57 of 102 agree	others agree 72%	watching
Answer spread	614 sampled answers compared with Z.AI's own endpoint. Very sensitive: it also picks up harmless setup differences, so it is shown for reference only	distance 0.34	for reference	info
Word probabilities	Top token probability at temperature 0, compared with the other providers	gap 0.32	returned	pass
Pinned routing	Request pinned to this endpoint with fallbacks off	served by this provider	served by this provider	pass
Reasoning switch	Asked for no reasoning on a one word answer	cannot be switched off	respected	pass

Provider reply

Reply as Wafer · fp4

No reply yet.

Results say whether a provider's answers match the maker's own endpoint. They are not a claim of intent. Quality checks are not running yet.