Treebeard 0.1.1 launcher refresh (pkg4): reasoning + speculation controls

run.sh/treebeard pkg4: TREEBEARD_REASONING=off|bounded|unrestricted and
TREEBEARD_SPECULATION=off|ngram|mtp|hybrid; PACKAGE.json optional_controls;
card + PACKAGE.md docs; SHA256SUMS re-pinned. Runtime bytes unchanged.

PACKAGE.json

@@ -3,7 +3,7 @@
   "name": "Treebeard-Qwen3.6-35B-A3B-GGUF",
   "product": "Treebeard",
   "version": "0.1.0-rc.3",
-  "packaging_revision": "pkg3",
+  "packaging_revision": "pkg4",
   "visibility": "public",
   "created_date": "2026-07-11",
   "publication": {
@@ -103,6 +103,31 @@
       "server_slots": 1
     }
   },
+  "optional_controls": {
+    "reasoning": {
+      "default": "off",
+      "modes": [
+        "off",
+        "bounded",
+        "unrestricted"
+      ],
+      "bounded_budget_tokens": {
+        "gpu": 64,
+        "cpu": 16
+      }
+    },
+    "speculation": {
+      "default": "off",
+      "modes": [
+        "off",
+        "ngram",
+        "mtp",
+        "hybrid"
+      ],
+      "native_mtp_layers": 1,
+      "validated_speedup": false
+    }
+  },
   "agent_benchmark": {
     "tool": "tool-eval-bench 2.1.0",
     "commit": "8b3259be7411fe27c7610d0de64ae1d3b622b9ef",

README.md

@@ -21,7 +21,7 @@ pipeline_tag: image-text-to-text
 # Treebeard
 
 Treebeard is a ready-to-run Linux package for Qwen3.6-35B-A3B. Release
-`0.1.0-rc.3` combines one 26.6 GB Q5_K_XL GGUF, the optional F16 vision
+`0.1.1` (runtime `0.1.0-rc.3`, launcher pkg4) combines one 26.6 GB Q5_K_XL GGUF, the optional F16 vision
 projector, official Qwen metadata and tokenizer files, platform runtimes, a
 one-command installer, an OpenAI-compatible server launcher, and the raw
 evidence behind its benchmark claims.
@@ -152,6 +152,9 @@ tokens. Override settings with environment variables:
 | `TREEBEARD_HOST` | `127.0.0.1` | API bind address |
 | `TREEBEARD_PORT` | `8093` | API port |
 | `TREEBEARD_MULTIMODAL` | `0` | Load the installed F16 projector |
+| `TREEBEARD_REASONING` | `off` | `off`, `bounded`, or `unrestricted` |
+| `TREEBEARD_REASONING_BUDGET` | `64` GPU / `16` CPU | Bounded thinking tokens |
+| `TREEBEARD_SPECULATION` | `off` | `off`, `ngram`, `mtp`, or `hybrid` |
 | `TREEBEARD_VERIFY` | `once` | `once`, `always`, or `never` |
 | `TREEBEARD_CPUSET` | unset | Optional `taskset` CPU list |
 
@@ -168,6 +171,16 @@ not the configuration behind the single-slot 94:
 TREEBEARD_PROFILE=throughput treebeard serve
 ```
 
+Reasoning is off by default, matching the validated 94/100 agent benchmark.
+`TREEBEARD_REASONING=bounded` enables a small thinking allowance (64 tokens
+on GPU, 16 on CPU; override with `TREEBEARD_REASONING_BUDGET`), and
+`unrestricted` removes the budget. API clients can still opt individual
+requests into thinking with request-level chat-template and thinking-budget
+controls. Speculative decoding is opt-in through
+`TREEBEARD_SPECULATION=ngram|mtp|hybrid` (`mtp` uses Qwen3.6's native
+one-layer MTP head carried by the GGUF); these modes ship in the 0.1.1
+launcher refresh but are not yet validated as a Treebeard speedup.
+
 ## Full repository package
 
 The Hugging Face repository is the complete package. After downloading it with

SHA256SUMS

@@ -2,15 +2,15 @@ cae3aa82fc25bb3f9f6ae4c826f8c08dea0bba30658c448a7f124737549ce345  .gitattributes
 50cbab8a892c5f2993b8c7351a99182507472def3b1374558308605d99b86b32  LICENSE
 94f29bbed6a22c35b992c5c6ebf0e7c92f13b836b90f36f461c9cf2f0f1d010d  LICENSE-RUNTIME
 2d04af27c51a8ba55b9f0cbfcf1df72270daf70cf5e5f6d6cbd2db794e9c2151  NOTICE.md
-1170075b0132584c68cd0f3d8d625e5c493a837ccb0d75adf4c9d5e70c4bc5de  PACKAGE.json
+7bdb30865738053188af49629d2fd9c0fbe9843d7c913727df01c7e554214f69  PACKAGE.json
 25233af7642e3a91bd52cc4aeefdbd4a117479088e06cf1aea5b6bedb443c506  Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf
-a2cddd57a1366568d96a3564f86e6f413cc649b7b815d8028033009203a52692  README.md
+a755e93db87862c16b9765b97a3bb9e2617d767a2aaa8565699e24a0257fa9aa  README.md
 e84f32a23fdda27689f868aa4a1a5621f41133e51a48d7f3efcbea2839574259  chat_template.jinja
 93a4693fa9d8392fbfccd4b3c9873f4bfdcb14fdede978b123d07d19675efe99  config.json
 c1b09db419119513247e9b8b912c4b9897106c9b20c6cada7e107d993c5435eb  configuration.json
 ae1c5fb00a620a316159929e229e4ca6a170efa6317e55164e614b120af417fd  docs/BENCHMARKS.md
 f648d15c3c111b72f775f9f08a43611d8398a9ff1855aba86d50576096a8b0c8  docs/NVIDIA.md
-b2338532cb94c426843d75bb9ad80b46f03bd788f6634da19c536579bb72e00b  docs/PACKAGE.md
+999acc5f1949ed80f5775503821b7f81ab1c31b47581e0ab884139dc49435398  docs/PACKAGE.md
 0daf18022e86b6e12510735aefa8ec52fea1c48b2f5c1136608287d48bda3f99  evidence/README.md
 efc91f319aef7a32c311a4d0f058dd87c92c1b96352fc322966c8ce2ed6fd4eb  evidence/agent/single-slot-94/LEGACY-RUN-IDENTITY.md
 6e3b2f8e317b123ae8fc35d91ca44a3765fdba1a8d50502be59624a6e24d3809  evidence/agent/single-slot-94/SHA256SUMS
@@ -143,14 +143,14 @@ a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d  merges.txt
 27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516  preprocessor_config.json
 bcbcf46eae0ada57cfa71645004fd0850781ef2bad7c9da57f7b922169baa421  profiles/quality.env
 97afe010ba8d3292456f875d115f03a257a95c12ae02d145c2bcaef449f4718e  profiles/throughput.env
-bae9586444cb35f2303f936cd0fffd47a845219e4bebd501a3133d45ffed1831  run.sh
+acd5c33ea206ba86814c5c739647a08f80b693332354c7a753a1eac9bade9206  run.sh
 48f8c4abb9467fcd0e4c0e8284251903f0e40b27dbf20786f3d9969478c0c81a  runtime/SHA256SUMS
 51497d7a72cfe8cdf3d66fcbf3ba36f8d5130084cc339cdc49995ec131126732  runtime/cpu-linux-x86_64/treebeard-0.1.0-rc.3-b9624-pkg3-cpu-ubuntu22.04-linux-x86_64.tar.gz
 0dfadb9ed53c66ab5a23b6731988d833b223af4e51c8f11ad9522a95eb0f04f3  runtime/cuda-linux-aarch64/treebeard-0.1.0-rc.3-b9624-pkg2-cuda13.3-linux-aarch64.tar.gz
 beec6bd316cc5285481d1545e66b31f984b1eb93e69fe05c7d0b1c5e1bafc756  runtime/sycl-linux-x86_64/treebeard-0.1.0-rc.3-b9624-pkg2-sycl-oneapi2026-linux-x86_64.tar.gz
 5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42  tokenizer.json
 5186f0defcd7f232382c7f0aebcd2252d073bb921ab240e407b7ae8745d2b29b  tokenizer_config.json
-e372e1c0faba070e61d6412607de6e6149cf0725b1c0a238f96aedbc61d5a3e0  treebeard
+16eb404820fb872f14460a546e79f3cf99c98d2314b40167c105ed08f2f1c132  treebeard
 3660ebd0bf0e9ce8c3a73af312077029793c3d9a8f52d973c7cd5fc9ca5e5f63  verify.sh
 7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13  video_preprocessor_config.json
 ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003  vocab.json

docs/PACKAGE.md

@@ -1,5 +1,11 @@
 # Package contract
 
+Model package: <https://huggingface.co/Frosty40/Treebeard-Qwen3.6-35B-A3B-GGUF>
+
+GitHub repository: <https://github.com/newjordan/treebeard>
+
+MoE algorithm explainer: <https://newjordan.github.io/treebeard/moe-routing.html>
+
 Treebeard pkg3 follows a standard Hugging Face GGUF layout. The complete GGUF
 model and official Qwen configuration and tokenizer files live at repository
 root. Runtimes, launch tools, and evidence are additive directories.
@@ -88,3 +94,40 @@ correctness where applicable, and a real server/API package smoke test.
 The model supports up to 262,144 context tokens, but actual usable context is
 bounded by available device or system memory. The launcher exposes overrides
 instead of claiming every host can sustain the maximum profile.
+
+## Reasoning and speculative decoding
+
+The validated profiles retain their existing context, slot, batch, and KV
+settings. Reasoning and speculation are independent, explicit controls layered
+on those resource profiles:
+
+| Setting | Behavior |
+| --- | --- |
+| `TREEBEARD_REASONING=off` | Default. Disables thinking in the server template while preserving explicit per-request overrides. |
+| `TREEBEARD_REASONING=bounded` | Enables thinking with a default 64-token GPU or 16-token CPU budget. |
+| `TREEBEARD_REASONING=unrestricted` | Enables thinking without a token budget. |
+| `TREEBEARD_SPECULATION=off` | Default. Leaves the runtime's no-speculation default unchanged. |
+| `TREEBEARD_SPECULATION=ngram` | Conservative `ngram-map-k` prompt-reuse drafting. |
+| `TREEBEARD_SPECULATION=mtp` | Conservative two-token drafting with the model's native MTP head. |
+| `TREEBEARD_SPECULATION=hybrid` | Tries n-gram drafting first, then native MTP. |
+
+Set `TREEBEARD_REASONING_BUDGET` to a positive integer to override the bounded
+default. The off setting is the server default; API clients can still opt an
+individual request into thinking with request-level chat-template and budget
+controls. Qwen3.6's native one-layer MTP head is carried by the GGUF, so `mtp`
+and `hybrid` do not require another model. Additional `llama-server` arguments
+may still be appended after `treebeard serve` for controlled experiments.
+
+On the pinned b9624 runtime, selective OpenAI-compatible thinking must set both
+`chat_template_kwargs.enable_thinking=true` and `thinking_budget_tokens=N` on
+the request while the launcher remains in its default `off` mode. The request's
+`max_tokens` limit includes both thought and answer tokens, and every model
+turn after a tool result starts with a fresh budget. Global bounded reasoning
+works on b9624, but overriding that global budget with a smaller request budget
+requires the newer request-precedence fix. The packaged RC3 binaries also
+predate newer Anthropic thinking-control translations; they require a runtime
+rebuild and are not claimed by this launcher-only change.
+
+The package makes no default speculative speed claim. N-gram hit rate, MTP
+acceptance, verification cost, memory pressure, and reasoning quality are
+workload-dependent and need matched evaluation before deployment.

run.sh

@@ -2,7 +2,7 @@
 set -Eeuo pipefail
 
 VERSION=0.1.0-rc.3
-PACKAGING_REVISION=pkg3
+PACKAGING_REVISION=pkg4
 MODEL_NAME=Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf
 MODEL_SHA256=25233af7642e3a91bd52cc4aeefdbd4a117479088e06cf1aea5b6bedb443c506
 MODEL_SIZE=26592508896
@@ -257,6 +257,15 @@ HOST="${TREEBEARD_HOST:-127.0.0.1}"
 PORT="${TREEBEARD_PORT:-8093}"
 MULTIMODAL="${TREEBEARD_MULTIMODAL:-0}"
 DRY_RUN="${TREEBEARD_DRY_RUN:-0}"
+REASONING="${TREEBEARD_REASONING:-off}"
+SPECULATION="${TREEBEARD_SPECULATION:-off}"
+
+if [[ "$BACKEND" == cpu ]]; then
+    DEFAULT_REASONING_BUDGET=16
+else
+    DEFAULT_REASONING_BUDGET=64
+fi
+REASONING_BUDGET="${TREEBEARD_REASONING_BUDGET:-$DEFAULT_REASONING_BUDGET}"
 
 require_positive_integer TREEBEARD_CONTEXT "$CONTEXT"
 require_positive_integer TREEBEARD_PARALLEL "$PARALLEL"
@@ -272,6 +281,25 @@ require_positive_integer TREEBEARD_PORT "$PORT"
 [[ "$FLASH_ATTN" == on || "$FLASH_ATTN" == off || "$FLASH_ATTN" == auto ]] ||
     fail "TREEBEARD_FLASH_ATTN must be on, off, or auto"
 
+case "$REASONING" in
+    off|unrestricted)
+        ;;
+    bounded)
+        require_positive_integer TREEBEARD_REASONING_BUDGET "$REASONING_BUDGET"
+        ;;
+    *)
+        fail "TREEBEARD_REASONING must be off, bounded, or unrestricted"
+        ;;
+esac
+
+case "$SPECULATION" in
+    off|ngram|mtp|hybrid)
+        ;;
+    *)
+        fail "TREEBEARD_SPECULATION must be off, ngram, mtp, or hybrid"
+        ;;
+esac
+
 mkdir -p "$CACHE_ROOT"
 chmod 700 "$CACHE_ROOT"
 verify_payload "$MODEL" "$MODEL_SHA256" "$MODEL_SIZE" "model"
@@ -325,6 +353,63 @@ args+=(
     -a "treebeard-$VERSION-Qwen3.6-35B-A3B-Q5-${BACKEND}-c${CONTEXT}-np${PARALLEL}"
 )
 
+case "$REASONING" in
+    off)
+        # Preserve the validated no-thinking server default. Clients can still
+        # opt individual requests into thinking with request-level controls.
+        args+=(--reasoning off --reasoning-budget -1)
+        REASONING_DETAIL=off
+        ;;
+    bounded)
+        args+=(--reasoning on --reasoning-budget "$REASONING_BUDGET")
+        REASONING_DETAIL="$REASONING_BUDGET"
+        ;;
+    unrestricted)
+        args+=(--reasoning on --reasoning-budget -1)
+        REASONING_DETAIL=unlimited
+        ;;
+esac
+
+case "$SPECULATION" in
+    off)
+        # The pinned runtime already defaults to one NONE sentinel. Do not add
+        # another --spec-type none entry on its append-style parser.
+        ;;
+    ngram)
+        # Short drafts and two required prompt hits are deliberately
+        # conservative; acceptance and speed remain workload-dependent.
+        args+=(
+            --spec-type ngram-map-k
+            --spec-ngram-map-k-size-n 12
+            --spec-ngram-map-k-size-m 16
+            --spec-ngram-map-k-min-hits 2
+        )
+        ;;
+    mtp|hybrid)
+        spec_types=draft-mtp
+        if [[ "$SPECULATION" == hybrid ]]; then
+            spec_types=ngram-map-k,draft-mtp
+            args+=(
+                --spec-ngram-map-k-size-n 12
+                --spec-ngram-map-k-size-m 16
+                --spec-ngram-map-k-min-hits 2
+            )
+        fi
+        # Qwen3.6's native one-layer MTP head is carried by the model GGUF.
+        # Keep the verification batch narrow until a workload proves a larger
+        # draft profitable on its backend and concurrency shape.
+        args+=(
+            --spec-type "$spec_types"
+            --spec-draft-n-max 2
+            --spec-draft-n-min 1
+            --spec-draft-p-min 0.20
+            --spec-draft-mtp-branch-k 1
+            --spec-draft-mtp-tree-width 1
+            --spec-draft-mtp-tree-depth 2
+        )
+        ;;
+esac
+
 if [[ "$BACKEND" == sycl ]]; then
     args+=(--no-op-offload)
 fi
@@ -333,8 +418,9 @@ if [[ "$MULTIMODAL" == 1 ]]; then
 fi
 args+=("$@")
 
-printf 'Treebeard %s %s: backend=%s profile=%s context=%s slots=%s multimodal=%s\n' \
-    "$VERSION" "$PACKAGING_REVISION" "$BACKEND" "$PROFILE" "$CONTEXT" "$PARALLEL" "$MULTIMODAL"
+printf 'Treebeard %s %s: backend=%s profile=%s context=%s slots=%s multimodal=%s reasoning=%s(%s) speculation=%s\n' \
+    "$VERSION" "$PACKAGING_REVISION" "$BACKEND" "$PROFILE" "$CONTEXT" "$PARALLEL" "$MULTIMODAL" \
+    "$REASONING" "$REASONING_DETAIL" "$SPECULATION"
 
 if [[ "$DRY_RUN" == 1 ]]; then
     printf 'Command:'

treebeard

@@ -22,6 +22,9 @@ Usage:
 Common settings:
   TREEBEARD_BACKEND=auto|cpu|sycl|cuda
   TREEBEARD_PROFILE=quality|throughput|custom
+  TREEBEARD_REASONING=off|bounded|unrestricted
+  TREEBEARD_REASONING_BUDGET=<tokens>  Bounded default: GPU 64, CPU 16
+  TREEBEARD_SPECULATION=off|ngram|mtp|hybrid
   TREEBEARD_CONTEXT=<tokens>
   TREEBEARD_HOST=127.0.0.1
   TREEBEARD_PORT=8093