The Pulse
Saturn Cloud offers operators compressed gpt-oss-120b, GLM-5.3 and DeepSeek-V4-Flash from Tileward
Operators using Saturn Cloud’s token factory can now deploy compressed versions of three open models from Tileward. Tileward reports that its compressed GLM-5.3 build is smaller than the publisher’s FP8 release and performed similarly on an MMLU test.

Operators using Saturn Cloud’s token factory can now deploy compressed versions of gpt-oss-120b, GLM-5.3 and DeepSeek-V4-Flash from Tileward. The partnership, described in a Saturn Cloud release on Nov. 20, makes the models available to operators running the platform on their own GPUs.
The offering is available to neoclouds, AI factory operators, telecommunications companies, sovereign AI providers and enterprises. Customers must contact either company to add the models to a deployment.
Swapping model weights
The token factory turns GPU fleets into managed inference and fine-tuning services that operators can sell under their own brands. It includes tenant isolation, quotas, metering and enterprise governance, and runs on NVIDIA Dynamo in public clouds, private clouds or on-premises.
Saturn Cloud says operators can replace published model weights with Tileward’s versions without changing their serving stack, tenants, rate cards or brand. The company says the memory freed by compression can allow operators to run more copies of a model, fit large models on fewer GPUs or serve more concurrent users with longer contexts.
The company also says serving more tokens from each GPU lowers the cost per token, leaving operators room to increase margins or reduce prices. Tileward offers compressed models that it says are up to 78% smaller than their published weights, and evaluates them against the originals using benchmarks including MMLU, SWE-bench Verified and AgentBench.
Compression and quality tests
In a separate evaluation of GLM-5.3, Tileward reports that its TWFP4 build is 464.8 GB, compared with 755.6 GB for Z.ai’s FP8 release and 1,506.7 GB for the 16-bit original. The compressed build scored 86.2% on MMLU, against 86.1% for the FP8 release and 86.3% for the 16-bit original.
The MMLU evaluation used 14,042 questions, five-shot prompting and no sampling. Tileward set a threshold before the run of no more than one point below the 16-bit original on the paired test.
On DeepSWE, the compressed build solved 86 of 113 tasks as a coding agent, compared with 74 for the FP8 release, in one run of each. Tileward cautions that a single run of each detects only large gaps and says it reads the result as evidence that its build holds up over long work, not that it is a better coder.
For Terminal-Bench 2.0, the compressed build solved 56 of 77 tasks, against 58 for the FP8 release, in the same comparison. Those figures are Tileward’s evaluation results, not a measurement of the models running in Saturn Cloud’s token factory.
Serving-stack benchmark
Separately, Saturn Cloud reports that it served Kimi K3 to 16 live agent sessions on NVIDIA GB300 GPUs at 7,380 total tokens per second per GPU and about 37 tokens per second per user. The company says the result matched the SemiAnalysis InferenceX benchmark recipe on the same GPUs; Tileward’s models run on the same serving stack.
Saturn Cloud’s website says both its run and the reference forced speculative-decoding acceptance. In the same test, a default Kimi K3 endpoint without speculative decoding delivered 5,557 total tokens per second per GPU at 22.4 tokens per second per user.
Sources
- Original source · Press release Saturn Cloud · Saturn Cloud brings Tileward’s compressed models to its tok… prnewswire.com
Other pages read
- Tileward tileward.com
- Saturn Cloud · page 2 saturn-cloud.co
In the tracker
Related stories
- FlashInfer on ROCm routes AMD GPU kernels automatically
- Qualcomm AI200 opens for select customer tests on Cirrascale
- Intel pitches D270 GPU for HPC with financial and AI benchmark claims
- Dynatrace expands observability in Gemini Enterprise workflows
- Faex adds investigational Ebola and H. pylori screening to AI platform