Qwen3.6 preserve_thinking chat-template flag and prompt-cache reuse

Facts

• Default template behaviour: reasoning is dropped from history except in the current tool-use turn. Between request N and N+1 the rendered prompt for turn N's assistant message therefore loses its `<think>...</think>` text. The tokens the server cached (prompt plus generated reasoning) no longer match the next prompt, so the shared prefix ends where the first dropped think block began. [source]
• Qwen's own explanation of the cache problem (maintainer reply on QwenLM/Qwen3.8 #131): stripping past thinking changes the prefix; a planned option to preserve thinking, even empty blocks in non-thinking mode, fixes cause 1; cause 2 is whitespace normalisation differences of thinking and tool-call blocks between model output and template re-render, which has no perfect fix because token-level generation and message-level templating are separate. [source]
• On hybrid models (Qwen3.5/3.6 delta-net layers) a mismatch is far costlier than on pure attention, because the recurrent state cannot be trimmed back to the divergence point; llama.cpp must restore a checkpoint at or before it or reprocess from zero. That is why the flag matters more here than for dense Qwen3. [source]
• Second, template-level cache bug: the stock template emits an empty `<think>\n\n</think>` wrapper on past assistant turns that had no reasoning (tool-call-only turns). Qwen3.8 #131 proposes guarding with `and reasoning_content`. LM Studio's froggeric MLX card calls this "empty preserve_thinking spam" and ships a rewritten template that emits think blocks only when reasoning is non-empty. [source]
• Client side contract: the server must receive reasoning back. llama-server returns it as `reasoning_content`; the client must echo it on the assistant message. If the client drops it, preserve_thinking has nothing to render. pi's `qwen-chat-template` thinkingFormat set only `enable_thinking`; the reported fix was to also send `preserve_thinking: true`. [source]
• llama.cpp naming split: the server-side generic flag `--reasoning-preserve` / `--no-reasoning-preserve` (env `LLAMA_ARG_REASONING_PRESERVE`) sets the template kwarg `preserve_reasoning`, not `preserve_thinking`. arg.cpp (master) injects `preserve_reasoning=true` by default after argument parsing if the user did not set it, and logs a deprecation warning when `preserve_reasoning` is set through `--chat-template-kwargs`. The Qwen template variable is `preserve_thinking`; a template that does not alias the two will not see the generic flag. [source]
• Template-side fix: froggeric's Qwen-Fixed-Chat-Templates added `preserve_reasoning` as an alias beside `preserve_thinking` (announced for v22, 2026-08-13), and its template treats an undefined `preserve_thinking` as true, so reasoning is preserved by default. [source]
• llama.cpp: `--chat-template-kwargs '{"preserve_thinking":true}'` (Unsloth and Qwen documented form); `--jinja` required; env `LLAMA_ARG_CHAT_TEMPLATE_KWARGS`; or per request in `chat_template_kwargs`. Router-mode `.ini` presets accept `chat-template-kwargs = {"preserve_thinking": true}`. Newer: `--reasoning-preserve` (see naming caveat). [source]
• Per request (any OpenAI-compatible server that forwards kwargs): `extra_body={"chat_template_kwargs": {"preserve_thinking": True}}` (Qwen model card). Alibaba Cloud Model Studio instead takes top-level `preserve_thinking: true`. [source]
• vLLM: `--default-chat-template-kwargs '{"preserve_thinking": true}'` (from Hermes #56004, held elsewhere). [source]
• LM Studio: no flag was found. The pi #3325 reporter used LM Studio with the stock Unsloth GGUF template and had to send the kwarg from the client; froggeric's MLX 8-bit build bakes a fixed template into the model instead. [source]
• oMLX: per-model settings in the admin panel include "chat template kwargs" (README), applied without restart; saved profiles can be exposed as `<model>:<profile>`. That is the place to set `{"preserve_thinking": true}`; no oMLX source was found naming the key. [source]
• Ollama: the library page (qwen3.6) advertises "Thinking Preservation" but no Ollama setting for it was found. [source]
• mlx_lm.server: no source found documenting chat_template_kwargs support or preserve_thinking. [source]
• Unsloth Studio: exposes "Think" and "Preserved Thinking" toggles for Qwen3.6 and Qwen3.8. [source]
• Agent harness example: OpenCode provider config with `"options": {"chat_template_kwargs": {"enable_thinking": true, "preserve_thinking": true}, "parallel_tool_calls": true}` (dzx.fr local coding guide). [source]
• 2026-04-09: Qwen3.8 #131 filed for empty historical think blocks; Qwen maintainer reply explains the two prefix-cache causes. [source]
• 2026-04-16 to 04-17: pi #3325 (LM Studio + Qwen3.6): tool calls loop with empty `{}` arguments after 2-3 turns until preserve_thinking is passed. [source]
• 2026-04-26: ggerganov on llama.cpp #22384 points to the model card. Reporter Tongas replies that with preserve_thinking=true turn 1 created no checkpoint (64-token floor) and turn 2 reprocessed; with the checkpoint patch turn 2 restored a checkpoint and processed 51 tokens. [source]
• 2026-05: ik_llama.cpp #1762 reporter notes mainline closed #22384 with "just enable preserve thinking" and asks whether that holds for ik_llama; a commenter lists the three prerequisites (client preserves thinking traces, requests pass `preserve_thinking: true` or `clear_thinking: false` for GLM-5.1, or set kwargs at server start). [source]
• 2026-07: llama.cpp adds generic `--reasoning-preserve` (default enabled, applies to templates with `supports_preserve_reasoning`); froggeric discussion #54 (2026-07-03) notes it does not map onto the stock Qwen key; froggeric adds the alias on 2026-08-13. [source]
• 2026-08: Qwen3.8 keeps the feature ("reasoning context from historical messages is retained via preserve_thinking"; llama-server logs "chat template supports preserving reasoning, consider enabling it via --reasoning-preserve" for Qwen3.8-27B). Unsloth Studio exposes the toggle for 3.8. [source]
• 2026-08-04: llama.cpp #25913 commenters list preserve_thinking or `--reasoning-preserve` as one of three things to check (with `--cache-ram` sizing and harness mutation such as `CLAUDE_CODE_ATTRIBUTION_HEADER=0`); one reporter replies it was not needed before a regression and would not fix it. [source]
• Setting the flag does not create a checkpoint the server never made. On hybrid models the first-turn checkpoint can be skipped (64-token minimum) and a repeated prompt can still reprocess; llama.cpp #22746 shows a healthy run with the flag (`restored context checkpoint ... n_past = 8887`, 400 tokens evaluated at 168 t/s of a 9,287-token prompt) next to a bad request later in the same log. [source]
• Qwen3.6-27B Q3_K_M run on llama.cpp #22746 used `--chat-template-kwargs '{"enable_thinking": true, "preserve_thinking": true}'` plus `--cache-ram 18432 --ctx-checkpoints 64 --checkpoint-every-n-tokens 8192 --parallel 2`; its log shows "reasoning-budget: re-activated on new start tag" repeatedly, so multiple think blocks per turn are normal under preserve_thinking. [source]
• Token cost: Unsloth says preserving thinking increases tokens used; Qwen's card says it can reduce overall tokens by avoiding redundant re-reasoning and improves KV cache utilisation. Qwen advises at least 128K context to keep thinking capability. [source]
• Interleaved-thinking rule from Qwen (#131 reply): in a tool chain after the last user message, think blocks (even empty) are kept; a message history of user, assistant, tool, assistant with no final tool call in the latest turn does not preserve historical blocks under the default template. [source]
• Disabling thinking (`enable_thinking: false` / `-rea off`) does not avoid churn: particula reports the assistant turn still emits two empty thinking tokens that are stripped next turn. [source]
• Client think-tag stripping: clients that do not echo `reasoning_content` (Hermes strips it for non-DeepSeek/Kimi/MiMo providers, held elsewhere) defeat the flag. A DeepSeek-style space pad would break prefix caching. [source]
• Parent-level mlx-lm interaction (not preserve_thinking-specific): mlx_lm.server tokenises prompts without prior reasoning, so cached key P+R+A+EOS diverges from next prompt P+A (mlx-lm #903, held in kv-cache-and-long-context-on-mac.md). No source shows whether passing preserve_thinking via mlx-lm fixes this. [source]
• Is `preserve_thinking` the fix for hybrid reprocessing? [source]
• For: ggerganov (llama.cpp #22384, 2026-04-26): "nothing to fix in llama-server"; Qwen model card says preserving thinking improves KV cache utilisation; Qwen's #131 reply says stripped traces alter the prefix; froggeric template notes claim 100% prefix KV hit when preserved (claim made in template documentation, unmeasured). [source]
• Against: Tongas (#22384) and #23030 reporter reproduced full reprocessing with the flag on; ik_llama.cpp #1762 thread questions it; a #25913 reporter says it "wouldn't fix the problem" after a regression; OpenCode #19081 commenter saw checkpoints erased with the flag set (held elsewhere). [source]
• Resolution suggested by the evidence: the flag removes one cause (template-dropped think text) only; checkpoint creation thresholds, client mutation and slot eviction remain independent causes. The two sides address different causes. [source]
• Whether the generic `--reasoning-preserve` suffices for Qwen: arg.cpp sets `preserve_reasoning` (default true); the Qwen template reads `preserve_thinking`. Commenters on froggeric #54 say the two names are different keys and were unaliased in the stock template; a repo-hosted alias exists only in the froggeric template. No source tested stock Qwen3.6 GGUF with only `--reasoning-preserve`. [source]
• Does stock Qwen3.6 GGUF template honour `preserve_reasoning`, or is `preserve_thinking` still required? Not tested in any source found. [source]
• Does mlx_lm.server (any release) forward `chat_template_kwargs`? Does oMLX apply its per-model kwargs before computing the cache key? [source]
• Is there any Ollama Modelfile or API control for preserve_thinking? [source]
• Is there a published before/after measurement of cache-hit tokens with the flag on a Mac (Metal) build? [source]
• Does the froggeric "100% prefix KV cache" claim survive whitespace normalisation of tool-call blocks (Qwen #131 cause 2)? [source]
• llama.cpp: pass `--jinja --chat-template-kwargs '{"enable_thinking": true, "preserve_thinking": true}'`; also pass `--reasoning-preserve`; use `--parallel 1`, raise `--ctx-checkpoints` and `--cache-ram`, lower the checkpoint interval (the #22746 reporter used 8192 and 64 checkpoints at about 150 MiB each); keep client prefix stable (`CLAUDE_CODE_ATTRIBUTION_HEADER=0`). [source]
• Ensure the client echoes `reasoning_content` back; verify with the server log (restored checkpoint line, `n_past` close to prompt length) rather than assuming. [source]
• oMLX: set `{"preserve_thinking": true}` in per-model chat template kwargs; keep SSD cache on. [source]
• Do not rely on the stock LM Studio GGUF template to receive the flag unless the client sends it. [source]
• Qwen3.6 keeps only the thinking from the latest user turn by default and was trained to use historical thinking when preserve_thinking is set [source]
• The model card says preserve_thinking can reduce overall token use in agent scenarios and can improve KV cache utilisation in thinking and non-thinking modes [source]
• The model card recommends at least 128K context to preserve thinking capabilities [source]
• Per request the kwarg is `chat_template_kwargs: {"preserve_thinking": true}` in `extra_body`; Alibaba Cloud Model Studio uses a top-level `preserve_thinking` [source]
• Unsloth documents `--chat-template-kwargs '{"preserve_thinking":true}'` for llama.cpp and says it increases tokens used but could raise accuracy in continued conversations [source]
• Unsloth Studio has Think and Preserved Thinking toggles for Qwen3.6 and Qwen3.8 [source]
• A Qwen maintainer reply on Qwen3.8 #131 names two prefix-cache causes: stripped past thinking, and whitespace normalisation differences in thinking and tool-call blocks [source]
• The same reply says the interleaved-thinking rule keeps think blocks (even empty) in the active tool-call chain after the last user message [source]
• Qwen3.8 #131 reports the stock template emits empty historical think blocks and proposes guarding on `reasoning_content` [source]
• LM Studio's froggeric Qwen3.6-27B MLX 8-bit build ships a rewritten template that emits think blocks only with real reasoning content and adds a `<|think_on|>`/`<|think_off|>` toggle [source]
• pi #3325: under LM Studio with `thinkingFormat: "qwen-chat-template"` Qwen3.6 emitted tool calls with empty `{}` arguments after 2-3 turns; the cause given is missing `preserve_thinking`, and the proposed fix sets `chat_template_kwargs {enable_thinking, preserve_thinking: true}` [source]
• llama.cpp master defines `--reasoning-preserve` / `--no-reasoning-preserve` (env LLAMA_ARG_REASONING_PRESERVE), default enabled, for templates with `supports_preserve_reasoning` [source]
• llama.cpp arg.cpp writes the key `preserve_reasoning` (not `preserve_thinking`) into default template kwargs and sets it to true when unspecified [source]
• llama.cpp logs a deprecation warning when `preserve_reasoning` or `enable_thinking` is passed through `--chat-template-kwargs` [source]
• A commenter on froggeric #54 says `preserve_thinking` is only a template argument and `--reasoning-preserve` is a generic server flag meant to work on all templates that implement it [source]
• froggeric (2026-08-13) added `preserve_reasoning` as an alias of `preserve_thinking` in template v22 [source]
• The froggeric template treats an undefined `preserve_thinking` as true [source]
• The froggeric template documentation claims preserving all past think blocks gives a 100% prefix KV cache hit rate on local engines (documentation claim, no measurement shown) [source]
• A llama.cpp #22384 reporter measured, with preserve_thinking=true and no checkpoint patch: turn 1 no checkpoint, turn 2 "forcing full prompt re-processing" of 75 tokens; with the patch: turn 2 restored a checkpoint and processed 51 tokens [source]
• On llama.cpp #22746 (Qwen3.6-27B Q3_K_M, preserve_thinking true, `--ctx-checkpoints 64`, `--checkpoint-every-n-tokens 8192`, `--cache-ram 18432`) a turn restored a 149.626 MiB checkpoint at n_past 8887 and processed 400 of 9,287 tokens at 168.49 t/s [source]
• Under preserve_thinking the same log shows llama.cpp's reasoning budget re-activating on each new start tag within one request [source]
• ik_llama.cpp #1762 lists three prerequisites for preserved thinking: client keeps thinking traces, requests pass `preserve_thinking: true` (or `clear_thinking: false` for GLM-5.1), or server-start kwargs [source]
• llama.cpp #25913 commenters list preserve_thinking or `--reasoning-preserve` as one check among `--cache-ram` sizing and harness mutation; one reporter says it would not fix their regression [source]
• particula reports `-rea off` still emits two empty thinking tokens per assistant turn that are stripped next turn, so it does not avoid context churn [source]
• particula recommends `--reasoning-preserve` as the real flag for templates advertising `supports_preserve_reasoning`, with `--parallel 1` and a raised checkpoint count (~150 MiB per checkpoint per slot) [source]
• llama-server logs "chat template supports preserving reasoning, consider enabling it via --reasoning-preserve" for Qwen3.8-27B [source]
• Qwen3.8 retains `preserve_thinking` and adds `reasoning_effort` [source]
• llama.cpp router-mode preset files accept `chat-template-kwargs = {"preserve_thinking": true}` [source]
• OpenCode provider options can carry `chat_template_kwargs: {enable_thinking: true, preserve_thinking: true}` [source]
• oMLX per-model settings include chat template kwargs, applied immediately without restart, and profiles exposable as `<model>:<profile>` [source]
• Ollama's Qwen3.6 library page lists "Thinking Preservation" as a model feature but shows no setting for it [source]
• A llama.cpp PR benchmark run of Qwen3.6 Q8_0 with MTP speculative decoding used preserve_thinking true throughout [source]
• No source found documents mlx_lm.server, Ollama or LM Studio's own UI exposing preserve_thinking; the mlx-lm and LM Studio gaps are an absence in the sources read [source]
• The flag removes only one of several independent causes of hybrid reprocessing (template-dropped reasoning), so it is necessary but not sufficient for cache reuse on Qwen3.5/3.6 [source]