---

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

Treebeard is a ready-to-run Linux package for Qwen3.6-35B-A3B. Release

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.

The model weights are one GGUF file. A runtime is still required to execute

that file, so the installer automatically adds the smallest compatible

runtime instead of pretending one binary can cover every CPU and GPU.

Install on Linux

This downloads about 26.7 GB for text use and installs treebeard under

~/.local/bin:

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

Then use the local OpenAI-compatible endpoint:

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}
  }'

Install vision support too:

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

The installer supports resumable downloads, verifies every installed file,

and safely reuses an already verified model. Useful options include

--backend, --install-dir, --bin-dir, and --allow-low-memory.

Requirements and supported paths

• Linux only for this release;
• 32 GB system or unified memory minimum, 40 GB recommended;
• roughly 29 GB free disk for text, plus 0.9 GB for vision;
• curl, tar, and GNU sha256sum for the installer;
• glibc 2.35 or newer for the portable CPU runtime.

| Selected backend | Platform | Host dependency | Validation |

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

| CPU | Linux x86_64 | None beyond glibc 2.35+ | Ryzen 9 5950X package, chat, and tool-call smoke |

| Intel SYCL | Linux x86_64 | Intel oneAPI 2026 and Level Zero | Arc Pro B70 correctness, performance, package, and agent gates |

| NVIDIA CUDA | Linux ARM64 | CUDA 13 runtime, cuBLAS, and compatible driver | GB10 correctness, performance, package, vision, tool-call, and agent gates |

NVIDIA acceleration in RC3 is the tested ARM64 GB10 path. An x86_64 machine

with an NVIDIA GPU remains usable through the portable CPU runtime, but its GPU

is not accelerated by this package yet. Hardware not listed above is

unverified even if the runtime starts.

94/100 agent result

The complete 69-scenario agent and tool-use suite reproduced the same result

on Intel and NVIDIA with one server slot:

| Result | Value |

| --- | ---: |

| Final score | 94 / 100 - Excellent |

| Points | 130 / 138 |

| Outcomes | 63 pass / 4 partial / 2 fail |

| Completed | 69 / 69 |

| Request errors | 0 |

| Total context | 262,144 tokens |

| Backends reproduced | Intel SYCL and NVIDIA CUDA |

View the standalone dark-mode

Treebeard benchmark report, or inspect

the packaged raw result and

NVIDIA replica. The two

backends produced the same verdict on every scenario.

The benchmark used tool-eval-bench 2.1.0 at commit 8b3259b, seed 42,

temperature 0, thinking disabled, one server slot, and one benchmark worker.

Deterministic mock tools were used; this is an agent-quality evaluation, not a

claim that arbitrary real-world actions are safe.

Performance highlights

| Measurement | Result |

| --- | ---: |

| NVIDIA GB10 native pp4096 | 2,422.325 tok/s |

| NVIDIA GB10 native tg128 | 59.614 tok/s |

| NVIDIA Blackwell Q8_0 direct 12-column speedup | 31.49% |

| NVIDIA Blackwell Q8_0 MoE down speedup | 4.01% |

| Intel B70 12-slot aggregate serving | 194.023 tok/s |

| CPU package chat generation, Ryzen 5950X | 9.30 tok/s |

| CPU package tool-call generation, Ryzen 5950X | 7.40 tok/s |

The CPU numbers are functional smoke measurements at 4,096 context, not a

cross-hardware performance promise. Exact commands, raw data, hashes, and

scope are in docs/BENCHMARKS.md.

Commands

treebeard serve       # start the API; this is also the default command
treebeard doctor      # show platform selection and the exact launch command
treebeard verify      # verify all installed files
treebeard status      # query the local health endpoint
treebeard report      # print the public benchmark URL
treebeard help

Quality mode is the default. It uses one slot and 262,144 context tokens on a

validated GPU backend; the portable CPU default is one slot and 32,768 context

tokens. Override settings with environment variables:

| Variable | Default | Meaning |

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

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

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

| TREEBEARD_CONTEXT | backend/profile value | Total context tokens |

| TREEBEARD_PARALLEL | backend/profile value | Fixed server slots |

| 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 |

For example:

TREEBEARD_CONTEXT=8192 TREEBEARD_PORT=8080 treebeard serve

The GPU throughput profile uses 12 slots. It is a saturation-serving profile,

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

Git LFS or the Hugging Face client, no second model repository is needed:

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

Using bash explicitly also works with archive and download clients that do

not preserve executable permission bits.

Package contents include:

• Qwen3.6-35B-A3B-UD-Q5_K_XL.gguf, the complete text model;
• mmproj-F16.gguf, the optional vision projector;
• standard Qwen config, tokenizer, processor, and chat-template files;
• portable CPU, Intel SYCL, and NVIDIA CUDA runtime archives;
• install.sh, treebeard, run.sh, profiles, and verification manifests;
• agent, CPU, SYCL, and NVIDIA evidence with reproduction scripts.

Every distributed file except the root manifest itself is covered by

SHA256SUMS. Runtime archives have their own internal file manifests and

provenance records. See docs/PACKAGE.md for the package

contract.

Multimodal mode

The validated co-resident NVIDIA profile uses 32,768 context tokens:

TREEBEARD_MULTIMODAL=1 TREEBEARD_CONTEXT=32768 treebeard serve

Multimodal mode uses additional device or system memory. Text and tool-use mode

remains the default because it is the profile covered by the 94/100 result.

Source and provenance

• Treebeard integration commit: 6a6dc2def952fe5e9b2da81e638968653b6be3db;
• release baseline: 3799687213eddbdc1389994f110fac0fa01f3e36;
• NVIDIA source patch SHA-256: c1e0780c96432059ea7a517f6ab2db935f1083da065ed0a9009a00d944c3415f;
• base model: Qwen/Qwen3.6-35B-A3B at metadata revision 995ad96e;
• GGUF source: unsloth/Qwen3.6-35B-A3B-GGUF at revision a483e9e6.

The source patch is distributed under evidence/nvidia/source. Historical raw

benchmark files retain their original aliases so their published hashes remain

verifiable; all current package and product surfaces use the Treebeard name.

Security and license

The API binds to loopback by default. Binding to a non-loopback address exposes

an unauthenticated inference service unless you add an authenticated TLS proxy

and firewall rules. Production tool execution should validate arguments,

authorize privileged actions, confirm external side effects, sandbox tools,

and retain audit logs.

The model and tokenizer are Apache-2.0. The llama.cpp-derived runtimes are MIT.

See LICENSE, LICENSE-RUNTIME, and NOTICE.md. Treebeard is not an official

Qwen, Unsloth, NVIDIA, Intel, Hugging Face, or llama.cpp release.