Fix: Adjust Vultr Model Context and Output Limits Based on Empirical Testing

[cyberofficial]

This pull request corrects the [limit] configurations for five models available on the Vultr Inference API to reflect their true, empirically tested context windows. The previous configurations contained aspirational limits that would lead to runtime errors. Due to limited documentation, the previous numbers were outdated.

Methodology:

I conducted a series of automated tests against the Vultr Inference API to determine the precise, stable context window for each model. The testing script rapidly found the upper bound, then used a guided search to fine-tune and verify the maximum token limit. The new configurations are based on these stable results.

Test Results Summary:

Model Name	Verified Avg. Limit	Safe Limit (Floor)	Hard Limit (Ceiling)
deepseek-r1-distill-qwen-32b	130,466 tokens	130,000 tokens	131,000 tokens
qwen2.5-coder-32b-instruct	15,940 tokens	15,000 tokens	16,000 tokens
deepseek-r1-distill-llama-70b	130,466 tokens	130,000 tokens	131,000 tokens
gpt-oss-120b	130,530 tokens	130,000 tokens	131,000 tokens
kimi-k2-instruct	63,667 tokens	63,000 tokens	64,000 tokens

Proposed Changes:

The [limit] section for each model file has been updated to be friendly for coding tasks (maximizing input context while reserving a generous output buffer) and to stay within the verified total token limits.

---

deepseek-r1-distill-qwen-32b.toml & deepseek-r1-distill-llama-70b.toml

 [limit]
-context = 128_000
-output = 32_768
+# VERIFIED TOTAL: ~130k. Reserving 8k for output.
+context = 121_808
+output = 8_192

---

gpt-oss-120b.toml

 [limit]
-context = 128_000
-output = 131_072
+# VERIFIED TOTAL: ~130k. Reserving 8k for output.
+context = 121_808
+output = 8_192

---

kimi-k2-instruct.toml

 [limit]
-context = 128_000
-output = 16_384
+# VERIFIED TOTAL: ~63k. Reserving 4k for output.
+context = 58_904
+output = 4_096

---

qwen2.5-coder-32b-instruct.toml

 [limit]
-context = 128_000
-output = 32_768
+# VERIFIED TOTAL: ~15k. Reserving 2k for output.
+context = 12_952
+output = 2_048