---

base_model: Tuwhy/Olmo-3-7B-Instruct-OPSA-Code

tags:

• gguf
• llama.cpp

---

Olmo-3-7B-Instruct-OPSA-Code-GGUF

GGUF conversion of Tuwhy/Olmo-3-7B-Instruct-OPSA-Code.

Files

• Olmo-3-7B-Instruct-OPSA-Code-Q4_K_M.gguf — Q4_K_M — 4.16 GiB

SHA256: 759bf7be6777a7fa6e3a1833736dbab5eba8924c2a27953e9274dbcdde408035

Conversion scope

Text-model GGUF conversion.

Conversion note

No MTP/NextN mismatch was detected by the preflight checks.

Validation gate

Before upload this exact file passed a real llama-server /v1/chat/completions smoke gate using the source chat template path. The gate checks:

• model loads successfully
• system/user chat template separation
• exact-output compliance for strict smoke prompts
• generation ends with finish_reason=stop instead of hitting the token cap
• no pseudo-role continuation such as /user or /assistant
• basic arithmetic generation
• no raw chat control-token leakage
• no thinking-tag leakage when reasoning is disabled

Reasoning mode used for smoke: auto.

llama.cpp chat-template compatibility

The source checkpoint's bundled Jinja template uses features that the tested llama.cpp minja parser cannot parse. This repository therefore includes a conservative text-only fallback template as chat-template-text-only.jinja. It preserves ordinary system/user/assistant text chat, but does not claim source tool/function-calling template compatibility.

Serve with:

llama-server -m Olmo-3-7B-Instruct-OPSA-Code-Q4_K_M.gguf --chat-template-file chat-template-text-only.jinja

Executable code smoke

Passed 3/3 local executable Python tasks. All tested generations ended normally: True. The smoke tasks covered palindrome detection, balanced parentheses, and interval merging; generated code was compiled and run against unit tests.

Source

Source revision used for conversion: 97e4404caa2159e6ddddf23b993a6234b699367b.

Source HEAD checked immediately before publish: 97e4404caa2159e6ddddf23b993a6234b699367b.

If these revisions differ while the source safetensors fingerprints are unchanged, the difference is metadata-only and publishing remains allowed.

All model credit belongs to the original model authors.