Skip to content

NVMAI 3.4

Choose a tag to compare

View all tags
@Pummelchen Pummelchen released this 13 Aug 02:36
· 875 commits to main since this release
v3.4
8d9ad63

NVMAI 3.4

  • No artificial output-token cap. When a client omits max_output_tokens / max_tokens / max_completion_tokens (Codex and OpenCode always omit them), generation defaults to the session context window instead of the old 2,048/4,096 defaults - the model replies until it is done, bounded only by the available context.
  • Two launchers replace the per-quantization launch_*.sh scripts and tools/nvmai-cli.sh: tools/server_launcher.sh (server only, prints the OpenAI API setup when ready) and tools/cli_launcher.sh (any of Codex / Qwen Code / OpenCode). Both ask the same five questions and accept the same positional arguments.
  • Reasoning mode (NVMAI_THINKING_MODE=1): replies are preceded by an explicit <think> reasoning pass; off by default. Both launchers add a think/nothink choice.
  • 72-combo coding-CLI benchmark harness (benchmark/combos.sh): 3 CLIs x 2 models x 3 quantizations x 2 modes x 2 reasoning, fastest-first and resumable.
Assets 2
Loading