Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Merge dev: take the qwen36 chat-flavor dispatch (#1777), preserve_thinking (#1767) and the engine's graceful exit (#1734) under the logprobs work
monotophic0901ffdd1842
diff --cc c/openai_server.pyindex 57e8b53e,a7ccf782..1e95c3b2--- a/c/openai_server.py+++ b/c/openai_server.py@@@ -938,11 -713,11 +939,11 @@@ def _parse_arch_tool_calls(reply, toolsif ARCH == "kimi":if tool_reply is not None:_sideband_text, calls = parse_k3_tool_calls(tool_reply, tools)- return reply.strip(), calls- return parse_k3_tool_calls(reply, tools) # compatibility with pre-#1147 engines+ return reply.strip(), calls, None, None+ return parse_k3_tool_calls(reply, tools) + (None, None) # pre-#1147 engines- if ARCH == "qwen38":+ if chat_flavor() == "qwen38":- return parse_qwen38_tool_calls(reply, tools)- return parse_tool_calls(reply, tools)+ return parse_qwen38_tool_calls(reply, tools) + (None, None)+ return _parse_tool_calls(reply, tools, track_spans)def _tool_stream_markers():diff --cc c/tests/test_openai_server.pyindex 710a003c,e6af5a13..3620d937--- a/c/tests/test_openai_server.py+++ b/c/tests/test_openai_server.py@@@ -27,13 -25,12 +27,15 @@@ from openai_server import (APIError, APparse_arch_tool_calls, parse_k3_tool_calls, parse_qwen38_tool_calls,read_engine_turn, render_chat, render_chat_for_arch,render_chat_glm53, render_chat_inkling, render_chat_kimi,- render_chat_olmoe,- render_chat_qwen38, render_chat_v4, render_chat_dsv41, serve,+ render_chat_olmoe, render_chat_qwen,+ render_chat_qwen38, render_chat_v4, render_chat_dsv41,+ _dsv4_tool_calls, serve,resolve_generation_prompt, split_thinking_reply,+ detect_chat_flavor, qwen36_has_vision,- starts_in_reasoning,+ split_thinking_reply_spans, starts_in_reasoning,+ parse_tool_calls_spans, parse_arch_tool_calls_spans,+ THINK_OPEN, THINK_CLOSE,+ _compose_span_maps, _cut_span_map, _project_span,stop_policy, tune_child_env)