---

license: other

license_name: apache-2.0-model-and-mit-runtime

base_model:

• Qwen/Qwen3.6-35B-A3B

tags:

• gguf
• qwen3.6
• llama.cpp
• nvidia
• cuda
• sycl
• intel-gpu
• cpu
• multimodal
• tool-calling
• long-context

pipeline_tag: image-text-to-text

---

Treebeard

Linux package for Qwen3.6-35B-A3B: one Q5_K_XL GGUF (~26.6 GB), optional F16

vision projector, Qwen tokenizer/metadata, platform runtimes (portable CPU,

Intel SYCL, NVIDIA CUDA), installer, and local llama-server with

/v1/chat/completions on loopback by default.

Package line on the card: release 0.1.1 / runtime 0.1.0-rc.3 (launcher pkg4)

unless a newer archive is listed in this repo. Weights are commodity open Qwen

GGUF; this repo is packaging + runtime evidence.

Site (evidence tables): https://newjordan.github.io/treebeard/

GitHub: https://github.com/newjordan/treebeard

Install on Linux

About 26.7 GB download (resumable). Installer checks file SHA-256. Roughly 32 GB

system or unified memory.

curl -fsSL https://raw.githubusercontent.com/newjordan/treebeard/main/install.sh | bash
~/.local/bin/treebeard doctor
~/.local/bin/treebeard serve

curl http://127.0.0.1:8093/health

curl http://127.0.0.1:8093/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "treebeard",
    "messages": [{"role": "user", "content": "Write a haiku about old forests."}],
    "max_tokens": 96,
    "chat_template_kwargs": {"enable_thinking": false}
  }'

Vision projector (+0.9 GB):

curl -fsSL https://raw.githubusercontent.com/newjordan/treebeard/main/install.sh | \
  bash -s -- --multimodal

Useful installer flags: --backend, --install-dir, --bin-dir, --allow-low-memory.

Platforms

| Backend | Platform | Host dependency | Validation row |

| --- | --- | --- | --- |

| CPU | Linux x86_64 | glibc 2.35+ | Ryzen 9 5950X package chat + tool-call smoke |

| Intel SYCL | Linux x86_64 | oneAPI 2026, Level Zero | Arc Pro B70 agent + serving evidence |

| NVIDIA CUDA | Linux ARM64 | CUDA 13, cuBLAS, driver | GB10 agent + native-bench + package smoke |

NVIDIA GPU acceleration in this package line is the ARM64 GB10 path. x86_64 NVIDIA

hosts use the portable CPU runtime unless a later package adds a CUDA x86 build.

Agent Bench freeze (named run)

| Field | Value |

| --- | --- |

| Test | tool-eval-bench 2.1.0 public 69 |

| Score | 94/100 (130/138); 63 pass / 4 partial / 2 fail; 0 request errors |

| Shape | np=1 · c=262144 · temp 0 · thinking off · seed 42 · commit 8b3259b |

| Hardware | Intel Arc Pro B70 and NVIDIA GB10 |

| Artifacts | result.json, nvidia-result.json |

Report UI: https://newjordan.github.io/treebeard/

This is a checksum-pinned freeze for a named package/runtime. It is the score in

those result files, not a guarantee for every local install or quant.

Control A/B (2026-07-28, stock Q5, Arc Pro B70)

Clean upstream SYCL binary versus Treebeard package binary + package env.

Same weights: Q5_K_XL sha256 25233af7642e3a91bd52cc4aeefdbd4a117479088e06cf1aea5b6bedb443c506.

| Test | Shape | Control | Treebeard | Artifact |

| --- | --- | ---: | ---: | --- |

| tool-eval-bench 69 | np=1 · c=262144 · seed 42 | 91/100 (125/138) | 91/100 (126/138) | GitHub agent-bench-ab REPORT |

| ho-pack-v1.1 | np=1 · seed 42 | 42/46 | 42/46 | same dir heldout/ |

| sequential tg_p50 | np=1 · c=32768 · 5×2 prompts | 77.1 t/s | 88.9 t/s | single-agent-ab |

| 12-agent ABA p50/agent | np=12 · n_predict=96 | 6.88 | 26.33 | base-vs-package-aba |

Multi-slot p50 is concurrent capacity. Sequential tg is single-stream.

Full notes: RELEASE-20260728.md.

Other measurements

| Test | Result | Artifact |

| --- | ---: | --- |

| llama-bench pp4096 (GB10) | 2,422.325 tok/s | evidence/nvidia / package BENCHMARKS |

| llama-bench tg128 (GB10) | 59.614 tok/s | same |

| Q8_0 12-col latency (Blackwell) | 31.49% lower | docs/BENCHMARKS.md |

| Q8_0 MoE-down latency (Blackwell) | 4.01% lower | same |

| 12-slot aggregate serving (B70) | 194.023 tok/s | docs/BENCHMARKS.md |

| CPU package chat smoke (5950X, c=4096) | 9.30 tok/s | evidence/cpu-linux-x86_64 |

| CPU package tool-call smoke | 7.40 tok/s | same |

CPU numbers are smoke at 4,096 context, not a cross-hardware promise.

Detail and scope: docs/BENCHMARKS.md.

Commands

treebeard serve       # default: start local API
treebeard doctor      # platform selection + launch command
treebeard verify      # verify installed files
treebeard status      # health endpoint
treebeard report      # public report URL
treebeard help

| Variable | Default | Meaning |

| --- | --- | --- |

| TREEBEARD_PROFILE | quality | quality, throughput, or custom |

| TREEBEARD_BACKEND | auto | auto, cpu, sycl, or cuda |

| TREEBEARD_CONTEXT | backend/profile | Total context tokens |

| TREEBEARD_PARALLEL | backend/profile | Server slots |

| TREEBEARD_HOST | 127.0.0.1 | Bind address |

| TREEBEARD_PORT | 8093 | Port |

| TREEBEARD_MULTIMODAL | 0 | Load 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 |

Default GPU quality profile: 1 slot, 262144 context (shape used for the freeze).

Throughput profile: 12 slots (serving saturation). Reasoning defaults off.

TREEBEARD_SPECULATION is opt-in; no speed claim attached on this card.

TREEBEARD_CONTEXT=8192 TREEBEARD_PORT=8080 treebeard serve
TREEBEARD_PROFILE=throughput treebeard serve
TREEBEARD_MULTIMODAL=1 TREEBEARD_CONTEXT=32768 treebeard serve

Full package layout

This HF repo is the package (Git LFS for the GGUF). After clone/download:

bash ./verify.sh
bash ./treebeard doctor
bash ./treebeard serve

Contents include the Q5 GGUF, optional mmproj-F16.gguf, tokenizer/config,

CPU/SYCL/CUDA runtime archives, install.sh / treebeard / run.sh, profiles,

SHA256SUMS, and evidence under evidence/. See docs/PACKAGE.md.

Provenance (package pin)

• Integration commit (package line): 6a6dc2def952fe5e9b2da81e638968653b6be3db
• Release baseline: 3799687213eddbdc1389994f110fac0fa01f3e36
• NVIDIA patch SHA-256: c1e0780c96432059ea7a517f6ab2db935f1083da065ed0a9009a00d944c3415f
• Base model: Qwen/Qwen3.6-35B-A3B metadata revision 995ad96e
• GGUF source: unsloth/Qwen3.6-35B-A3B-GGUF revision a483e9e6

Security and license

API binds to loopback by default. Non-loopback bind is an unauthenticated

inference surface unless you add TLS auth and firewall rules.

Model/tokenizer: Apache-2.0. llama.cpp-derived runtimes: MIT. See LICENSE,

LICENSE-RUNTIME, NOTICE.md. Unofficial packaging; not an official Qwen,

Unsloth, NVIDIA, Intel, Hugging Face, or llama.cpp product.