I've been running Ornith 1.0 35B (Q4_K_M) on llama.cpp with the stock Qwen template for weeks, and just switched to froggeric's v21.3. The improvement is measurable and comes from several concrete fixes:
KV Cache hit rate → 100% — The stock template strips reasoning content from past turns, invalidating the prefix cache on every turn. Froggeric preserves thoughts chronologically (preserve_thinking: true), so multi-turn conversations reuse cached KV. On llama.cpp with --cache-reuse, this is the difference between 30 t/s and 65 t/s after a few turns.
No more "empty think" poisoning — The stock template replaces past thoughts with empty thinking\n\n response blocks + a directive to call a tool immediately. This creates a toxic pattern where the model associates empty thoughts with tools and full thoughts with "forbidden" text, causing an 80%+ premature stall rate. Froggeric abolished empty think injection entirely.
Two-tier agentic error escalation — Tracks a consecutive_failures counter. First failure: seeds reasoning at a different token position. Second failure: bypasses thinking and forces a corrected action. This fixes the "model keeps calling the same broken tool forever" problem.
AST flattening → ~80% throughput improvement on C++ engines (llama.cpp, LM Studio). The stock template's deeply nested Jinja creates severe parsing bottlenecks.
minijinja compatibility — The stock template uses Python-only Jinja2 features (| items, loop.previtem, | first) that crash or silently misbehave on C++ runtimes. Froggeric rewrote every filter to be 100% minijinja-safe.
Smart false-positive detection — Instead of broad "error" substring matching (which triggers false retry loops on successful DB returns containing the word "error"), uses strict structural guards: Exception:, "error":, Traceback, command not found.
Dynamic payload truncation — max_tool_arg_chars / max_tool_response_chars limiters prevent context window blowouts from massive API returns.
Inline thinking toggle — <|think_on|> / <|think_off|> anywhere in system/user prompts to control reasoning per-query. Huge for BI/analytics use cases.
Bottom line: Every fix addresses a specific, reproducible failure mode in the stock Qwen template — KV cache invalidation, agentic stalls, C++ parsing crashes, false error loops, and context window blowouts. If you're running Ornith (or any Qwen3.5/3.6 model) on llama.cpp, vLLM, or LM Studio, this template is essentially required for reliable agentic behavior.
One caveat: remove any chat_template_kwargs overrides you were using (enable_thinking, auto_disable_thinking_with_tools) — the new template handles everything internally via <|think_on|>/<|think_off|> tags. Just pass --chat-template-file chat_template.jinja and let it work.
