✨ Meet the model
Meet your local AI companion for ideas, code, images and conversation. Qwen3.6-35B v2 by oktayd brings these interests together in one downloadable model. Choose the edition that fits your setup, give it a task and shape its style with your own instructions.
Table with columns: Focus | Focus |
|---|
| 🧠 Learn & explore | Ask about science, work through a problem or get an explanation in everyday language. |
| 💻 Build & fix | Try a website idea, draft code or work through a bug together. |
| 🛠️ Connect your tools | Use it in a tool-enabled app for structured requests and workflows. Your app supplies and executes the tools. |
| 👁️ Bring an image | Ask about a screenshot, document, diagram or scene. |
| 🎭 Set the personality | Explore stories, roleplay and different conversational styles through your instructions. |
| 📦 Run it your way | Three editions and seven GGUF sizes, from compact experiments to higher-precision weights. |
These are things to try, not promises of perfect results. It can make mistakes or repeat itself; check important answers. Adult-oriented training is included (18+).
🧠 A team of experts inside one model
MoE means Mixture of Experts. Think of a team of specialists: for each token, a router chooses which expert networks should contribute. This model has about 35 billion parameters in total, with roughly 3 billion active per token; its configuration selects 8 of 256 routed experts.
That saves computation compared with activating every expert at once. It does not turn a 35B download into a 3B-sized model: the expert weights still need disk space and accessible RAM/VRAM. Quantized editions make local use more practical. The experts are learned networks, not separate installed apps or named profession-specific agents.
🕶️ Street knowledge. Business mind. Your style.
The personality direction is direct, sharp-witted and business-minded: a street-smart conversation partner with humor, creative confidence and room for disagreement. The training mix includes internet culture, slang, practical business topics and personality-oriented conversations. Think less formal textbook, more an opinionated partner for brainstorming, writing and exploring alternatives.
You set the tone: professional, casual, blunt or playful. The aim is personality without blind agreement. This is a style and training focus, not a guarantee of factual expertise or flawless judgment.
🚀 Where should I start?
Table with columns: Your setup, Choose| Your setup | Choose |
|---|
| I want a local chat app | Ollama edition and its guided importer |
| I use llama.cpp or a compatible GGUF app | GGUF / Llama edition; Q4_K_M is a starting point if it fits your memory |
| I work with Python or want the full weights | BF16 / FT edition |
Smaller files need less memory, but the lowest-bit versions can lose substantial quality. Vision needs the included image processor/projector as well as the language-model weights.
💬 Try asking
- “Explain this screenshot and suggest my next step.”
- “Turn this idea into a simple website, then explain how to run it.”
- “Find the bug in this function and show a corrected version.”
- “Be a witty, direct writing partner and help develop this character.”
📚 Training in brief
34,000 record uses in three runs: broad knowledge, STEM, coding, tools and conversation, with selected visual examples. The private packages cover synthetic 3D model adult learning and broad adult learning; a small Chinese-caption sample complements the visual mix. Dataset sources and training counts.
⚙️ Architecture & validation
Qwen3_5MoeForConditionalGeneration; approximately 35B total parameters, 256 experts, 8 selected per token (the previous A3B label).- Native vision-language architecture; separate projector required for GGUF image input. Text and elementary red/blue-image smoke tests were recorded. Video, 3D consistency and full desktop/browser agents are not certified by those tests.
- Training covered selected knowledge, instruction/agent, coding, personality and visual data: 34,000 record uses across 6,000 + 12,000 + 16,000. Record uses are not unique records. Training loss is not benchmark accuracy.
- All 19 source MTP tensors were preserved. MTP/speculative acceleration is not enabled or validated by these benchmark results. Backend support must be tested separately.
- Tool-call formatting and end-of-answer behavior are diagnostic targets, not guaranteed features. Known schema mistakes, factual errors and repetition remain.
🚀 Start on your device
Choose your hardware, then expand Ubuntu/Linux or Windows/PowerShell inside it. You only need one edition and one quantization. Installation examples are not new benchmark results.
For a machine with enough GPU memory. Validated on Ubuntu / NVIDIA H200 with Transformers 5.16.1. BF16 weights alone are about 70 GB, plus runtime memory. A Windows client does not provide the remote server's VRAM; this package does not fit a single RTX 5090 or an 8-GB laptop GPU.
Install Python 3.11, virtual-environment support and a CUDA-enabled PyTorch build from the official PyTorch selector that matches your NVIDIA driver. On a fresh Ubuntu machine:
sudo apt-get update
sudo apt-get install -y python3-venv python3-pip
mkdir -p qwen-v2-server
cd qwen-v2-server
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
# Install the CUDA-enabled PyTorch command from the official selector here.
python -m pip install transformers==5.16.1 accelerate pillow
python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"
If the GPU check fails, fix the driver/PyTorch installation before downloading the model. Save the shared Python example below as qwen_chat.py, then run python qwen_chat.py. The first run downloads the BF16 weights.
Native Windows — WSL2 is not required. Install Python 3.11 and a compatible NVIDIA driver, then use PowerShell 7. This native installation route is separate from the measured H200 validation, which was performed on Ubuntu:
New-Item -ItemType Directory -Path .\qwen-v2-server -Force | Out-Null
Set-Location .\qwen-v2-server
py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install --upgrade pip
# Run the Windows CUDA PyTorch install command from the official selector,
# using .\.venv\Scripts\python.exe -m pip instead of pip.
.\.venv\Scripts\python.exe -m pip install transformers==5.16.1 accelerate pillow
.\.venv\Scripts\python.exe -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"
Save the shared example below as qwen_chat.py, then execute .\.venv\Scripts\python.exe .\qwen_chat.py. Use only Windows GPU/driver combinations supported by PyTorch and with sufficient memory.
If the H200 belongs to a remote Ubuntu server, Windows is just the client. Connect using Windows OpenSSH (replace the example and use your actual port/key), then follow the Ubuntu instructions in that remote shell:
Optional alternative, not required for native Windows. Follow Microsoft's WSL installation guide. From an administrator PowerShell, only if you want this alternative:
Restart if requested and complete Ubuntu's first-run account setup. Then open Ubuntu:
Inside Ubuntu, follow this device's Ubuntu/Linux instructions and create a separate Linux Python environment. Keep the Windows NVIDIA driver current; use NVIDIA's CUDA-on-WSL guidance rather than installing a Linux display driver inside WSL. Verify GPU access before loading weights.
Use either the Windows or WSL model server during the test, not both on the same port. WSL does not add VRAM. Reuse model downloads where practical, but do not reuse Windows Python environments inside Linux; Ollama stores/imports may create another copy. Native Windows remains the default path above.
Shared Python example — use after either installation:
import torch
from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration
repo = "oktayd/Qwen3.6-35B-v2-MoE-Ablit-Heretic-Uncensor-Hermes-MTP-Vision-FT"
processor = AutoProcessor.from_pretrained(repo)
model = Qwen3_5MoeForConditionalGeneration.from_pretrained(
repo, dtype=torch.bfloat16, device_map="auto"
)
messages = [{"role": "user", "content": [
{"type": "text", "text": "Explain gravity briefly."}
]}]
inputs = processor.apply_chat_template(
messages, tokenize=True, add_generation_prompt=True,
return_dict=True, return_tensors="pt", enable_thinking=False
).to(model.device)
with torch.inference_mode():
output = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(processor.batch_decode(
output[:, inputs.input_ids.shape[1]:], skip_special_tokens=True
)[0])
For images, include an image content item supported by AutoProcessor. Large images and long contexts increase memory use.
Use the GGUF / Llama edition. Q4_K_M is about 20.22 GiB, plus the 0.84-GiB projector and runtime memory. RTX 5090 / Q4_K_M was benchmarked with llama.cpp commit 427291b5b34cd914a31b3fd3b61a68f6184f4b9f on Ubuntu. The Windows steps below are installation guidance, not a separate Windows 5090 result.
Install Python, the HF CLI and a CUDA-enabled llama.cpp build. Put llama-server on PATH. Choose a driver/toolkit/build supporting the RTX 5090; a CPU-only binary is not sufficient for GPU offload.
mkdir -p qwen-v2-q4
cd qwen-v2-q4
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade huggingface_hub
llama-server --list-devices
repo="oktayd/Qwen3.6-35B-v2-MoE-Ablit-Heretic-Uncensor-Hermes-MTP-Vision-Llama"
hf download "$repo" Qwen3.6-35B-v2-Q4_K_M.gguf mmproj-Qwen3.6-35B-v2-F16.gguf SHA256SUMS LICENSE --local-dir .
sha256sum --check --ignore-missing SHA256SUMS
llama-server -m Qwen3.6-35B-v2-Q4_K_M.gguf --mmproj mmproj-Qwen3.6-35B-v2-F16.gguf -ngl 99 -c 4096 -np 1 --host 127.0.0.1 --port 8080 --chat-template-kwargs '{"enable_thinking":false}'
Stop if checksums fail or the device listing does not show the intended NVIDIA GPU. In another terminal:
curl http://127.0.0.1:8080/v1/chat/completions -H "Content-Type: application/json" -d '{"messages":[{"role":"user","content":"Explain gravity briefly."}],"max_tokens":256,"temperature":0,"stream":false}'
Install Python 3.11 and download the matching Windows CUDA package and any required CUDA runtime DLL package from the official llama.cpp releases. Extract the full package, preserving its DLLs; add that folder to PATH for this terminal. Do not substitute a CPU-only build. These commands assume llama-server.exe is on PATH. Use PowerShell 7 for native JSON argument handling.
New-Item -ItemType Directory -Path .\qwen-v2-q4 -Force | Out-Null
Set-Location .\qwen-v2-q4
py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install --upgrade huggingface_hub
llama-server.exe --list-devices
$qwenRepo = "oktayd/Qwen3.6-35B-v2-MoE-Ablit-Heretic-Uncensor-Hermes-MTP-Vision-Llama"
.\.venv\Scripts\hf.exe download $qwenRepo Qwen3.6-35B-v2-Q4_K_M.gguf mmproj-Qwen3.6-35B-v2-F16.gguf SHA256SUMS LICENSE --local-dir .
if ($LASTEXITCODE -ne 0) { throw "Download failed" }
$qwenFiles = @("Qwen3.6-35B-v2-Q4_K_M.gguf", "mmproj-Qwen3.6-35B-v2-F16.gguf")
$qwenChecks = @{}
Get-Content .\SHA256SUMS | ForEach-Object {
if ($_ -match '^([0-9a-fA-F]{64}) (.+)$') { $qwenChecks[$matches[2]] = $matches[1] }
}
foreach ($qwenFile in $qwenFiles) {
if (-not $qwenChecks.ContainsKey($qwenFile)) { throw "Missing checksum: $qwenFile" }
if ((Get-FileHash -LiteralPath $qwenFile -Algorithm SHA256).Hash -ne $qwenChecks[$qwenFile]) { throw "Checksum mismatch: $qwenFile" }
}
llama-server.exe -m Qwen3.6-35B-v2-Q4_K_M.gguf --mmproj mmproj-Qwen3.6-35B-v2-F16.gguf -ngl 99 -c 4096 -np 1 --host 127.0.0.1 --port 8080 --chat-template-kwargs '{"enable_thinking":false}'
Confirm that --list-devices lists the RTX 5090. In another PowerShell window:
$qwenBody = @{
messages = @(@{ role = "user"; content = "Explain gravity briefly." })
max_tokens = 256
temperature = 0
stream = $false
} | ConvertTo-Json -Depth 5
$qwenReply = Invoke-RestMethod -Uri "http://127.0.0.1:8080/v1/chat/completions" -Method Post -ContentType "application/json" -Body $qwenBody -TimeoutSec 300
$qwenReply.choices[0].message.content
Optional alternative, not required for native Windows. Follow Microsoft's WSL installation guide. From an administrator PowerShell, only if you want this alternative:
Restart if requested and complete Ubuntu's first-run account setup. Then open Ubuntu:
Inside Ubuntu, follow this device's Ubuntu/Linux instructions and create a separate Linux Python environment. Keep the Windows NVIDIA driver current; use NVIDIA's CUDA-on-WSL guidance rather than installing a Linux display driver inside WSL. Verify GPU access before loading weights.
Use either the Windows or WSL model server during the test, not both on the same port. WSL does not add VRAM. Reuse model downloads where practical, but do not reuse Windows Python environments inside Linux; Ollama stores/imports may create another copy. Native Windows remains the default path above.
Keep the service on localhost. For a remote server, use SSH port forwarding rather than opening an unauthenticated public endpoint. Start with a short context; other GPU workloads reduce available VRAM.
IQ4_XS: about 17.86 GiB, plus a 0.84-GiB vision projector and runtime memory. It does not fit fully into 8 GB VRAM; GPU + CPU/RAM offload is required. Allow about 40 GB free disk for the download and a possible second imported copy. These are planning estimates, not measured peaks. Close memory-heavy apps and use one model at a time.
The complete IQ4_XS laptop comparison is still pending. The importer was previously tested with Ollama 0.33.3; installed versions and device-specific performance are recorded separately.
Install Ollama using its official Linux instructions, plus Python 3 and virtual-environment support. Start the Ollama service, or run ollama serve in a separate terminal if it is not already running. Do not start a second server on an occupied port.
mkdir -p qwen-v2-iq4
cd qwen-v2-iq4
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade huggingface_hub requests
repo="oktayd/Qwen3.6-35B-v2-MoE-Ablit-Heretic-Uncensor-Hermes-MTP-Vision-Ollama"
hf download "$repo" Qwen3.6-35B-v2-IQ4_XS.gguf mmproj-Qwen3.6-35B-v2-F16.gguf import_ollama.py SHA256SUMS LICENSE --local-dir .
python import_ollama.py --quant IQ4_XS
ollama run qwen3.6-35b-v2:iq4_xs --think=false
For a lighter, bounded first request instead of interactive chat:
curl http://127.0.0.1:11434/api/chat \
-H "Content-Type: application/json" \
-d '{"model":"qwen3.6-35b-v2:iq4_xs","messages":[{"role":"user","content":"Explain gravity briefly."}],"think":false,"stream":false,"options":{"num_ctx":2048,"num_predict":256}}'
Install Ollama for Windows and Python 3.11. Keep the Ollama app running. The commands use a dedicated Python environment and do not change PowerShell execution policy.
New-Item -ItemType Directory -Path .\qwen-v2-iq4 -Force | Out-Null
Set-Location .\qwen-v2-iq4
py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install --upgrade huggingface_hub requests
$qwenRepo = "oktayd/Qwen3.6-35B-v2-MoE-Ablit-Heretic-Uncensor-Hermes-MTP-Vision-Ollama"
.\.venv\Scripts\hf.exe download $qwenRepo Qwen3.6-35B-v2-IQ4_XS.gguf mmproj-Qwen3.6-35B-v2-F16.gguf import_ollama.py SHA256SUMS LICENSE --local-dir .
if ($LASTEXITCODE -ne 0) { throw "Download failed" }
.\.venv\Scripts\python.exe .\import_ollama.py --quant IQ4_XS
if ($LASTEXITCODE -ne 0) { throw "Import failed" }
ollama run qwen3.6-35b-v2:iq4_xs --think=false
For a lighter, bounded first request instead of interactive chat:
$qwenRequest = @{
model = "qwen3.6-35b-v2:iq4_xs"
messages = @(@{ role = "user"; content = "Explain gravity briefly." })
think = $false
stream = $false
options = @{ num_ctx = 2048; num_predict = 256 }
} | ConvertTo-Json -Depth 5
$qwenReply = Invoke-RestMethod -Uri "http://127.0.0.1:11434/api/chat" -Method Post -ContentType "application/json" -Body $qwenRequest -TimeoutSec 300
$qwenReply.message.content
Optional alternative, not required for native Windows. Follow Microsoft's WSL installation guide. From an administrator PowerShell, only if you want this alternative:
Restart if requested and complete Ubuntu's first-run account setup. Then open Ubuntu:
Inside Ubuntu, follow this device's Ubuntu/Linux instructions and create a separate Linux Python environment. Keep the Windows NVIDIA driver current; use NVIDIA's CUDA-on-WSL guidance rather than installing a Linux display driver inside WSL. Verify GPU access before loading weights.
Use either the Windows or WSL model server during the test, not both on the same port. WSL does not add VRAM. Reuse model downloads where practical, but do not reuse Windows Python environments inside Linux; Ollama stores/imports may create another copy. Native Windows remains the default path above.
The importer verifies the selected weight/projector hashes and includes both. Its interactive defaults are 4,096 context tokens and 1,024 output tokens; the API examples use 2,048 / 256. Check CPU/GPU allocation with ollama ps. No Transformers/BF16 download is needed. Ollama chat API.
IQ2_M: about 11.70 GiB, plus a 0.84-GiB vision projector and runtime memory. It does not fit fully into 8 GB VRAM; GPU + CPU/RAM offload is required. Allow about 27 GB free disk for the download and a possible second imported copy. These are planning estimates, not measured peaks. Close memory-heavy apps and use one model at a time.
Experimental low-bit option. This is IQ2_M, not Q2_K. Reasoning, code and format accuracy can suffer. Initial small Windows smoke checks have started; a full laptop quality/performance comparison is still pending.
Install Ollama using its official Linux instructions, plus Python 3 and virtual-environment support. Start the Ollama service, or run ollama serve in a separate terminal if it is not already running. Do not start a second server on an occupied port.
mkdir -p qwen-v2-iq2
cd qwen-v2-iq2
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade huggingface_hub requests
repo="oktayd/Qwen3.6-35B-v2-MoE-Ablit-Heretic-Uncensor-Hermes-MTP-Vision-Ollama"
hf download "$repo" Qwen3.6-35B-v2-IQ2_M.gguf mmproj-Qwen3.6-35B-v2-F16.gguf import_ollama.py SHA256SUMS LICENSE --local-dir .
python import_ollama.py --quant IQ2_M
ollama run qwen3.6-35b-v2:iq2_m --think=false
For a lighter, bounded first request instead of interactive chat:
curl http://127.0.0.1:11434/api/chat \
-H "Content-Type: application/json" \
-d '{"model":"qwen3.6-35b-v2:iq2_m","messages":[{"role":"user","content":"Explain gravity briefly."}],"think":false,"stream":false,"options":{"num_ctx":2048,"num_predict":256}}'
Install Ollama for Windows and Python 3.11. Keep the Ollama app running. The commands use a dedicated Python environment and do not change PowerShell execution policy.
New-Item -ItemType Directory -Path .\qwen-v2-iq2 -Force | Out-Null
Set-Location .\qwen-v2-iq2
py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install --upgrade huggingface_hub requests
$qwenRepo = "oktayd/Qwen3.6-35B-v2-MoE-Ablit-Heretic-Uncensor-Hermes-MTP-Vision-Ollama"
.\.venv\Scripts\hf.exe download $qwenRepo Qwen3.6-35B-v2-IQ2_M.gguf mmproj-Qwen3.6-35B-v2-F16.gguf import_ollama.py SHA256SUMS LICENSE --local-dir .
if ($LASTEXITCODE -ne 0) { throw "Download failed" }
.\.venv\Scripts\python.exe .\import_ollama.py --quant IQ2_M
if ($LASTEXITCODE -ne 0) { throw "Import failed" }
ollama run qwen3.6-35b-v2:iq2_m --think=false
For a lighter, bounded first request instead of interactive chat:
$qwenRequest = @{
model = "qwen3.6-35b-v2:iq2_m"
messages = @(@{ role = "user"; content = "Explain gravity briefly." })
think = $false
stream = $false
options = @{ num_ctx = 2048; num_predict = 256 }
} | ConvertTo-Json -Depth 5
$qwenReply = Invoke-RestMethod -Uri "http://127.0.0.1:11434/api/chat" -Method Post -ContentType "application/json" -Body $qwenRequest -TimeoutSec 300
$qwenReply.message.content
Optional alternative, not required for native Windows. Follow Microsoft's WSL installation guide. From an administrator PowerShell, only if you want this alternative:
Restart if requested and complete Ubuntu's first-run account setup. Then open Ubuntu:
Inside Ubuntu, follow this device's Ubuntu/Linux instructions and create a separate Linux Python environment. Keep the Windows NVIDIA driver current; use NVIDIA's CUDA-on-WSL guidance rather than installing a Linux display driver inside WSL. Verify GPU access before loading weights.
Use either the Windows or WSL model server during the test, not both on the same port. WSL does not add VRAM. Reuse model downloads where practical, but do not reuse Windows Python environments inside Linux; Ollama stores/imports may create another copy. Native Windows remains the default path above.
The importer verifies the selected weight/projector hashes and includes both. Its interactive defaults are 4,096 context tokens and 1,024 output tokens; the API examples use 2,048 / 256. Check CPU/GPU allocation with ollama ps. No Transformers/BF16 download is needed. Ollama chat API.
Validation and provenance
RELEASE-VALIDATION.json records the pinned base revision, final adapter checksum,
tensor coverage, smoke outputs, and exact pre-save/post-reload logit comparison.
BF16 merging introduces rounding relative to unmerged LoRA inference; this
difference is recorded rather than claimed to be bitwise identical.
SHA256SUMS covers the files in this release. The model inherits the source's
Apache-2.0 license. This is a research fine-tune; no claim of universal benchmark
improvement or removal of memorization is made.
Bounded inference (runtime safeguard)
The release enables KV caching for inference. This corrects a training-time
use_cache: false export setting; it does not repair learned reasoning behavior.
Short diagnostic tests observed repetitive generations, particularly with thinking
enabled. enable_thinking=False is the recommended starting point for interactive
use, not a guarantee of correctness. Thinking mode remains available explicitly.
The optional, inspectable q36_runtime.py helper supports batch size one:
from q36_runtime import generate_bounded
result = generate_bounded(model, processor.tokenizer, inputs,
max_new_tokens=1024, max_seconds=45)
print(result['text'])
print(result['completed'], result['finish_reason'])
Build inputs with enable_thinking=False as above. The helper stops on EOS,
token/time limits, or a long repeated-token-block heuristic. A limit/loop stop is
reported as incomplete, not as a successful answer. It never executes tools
or retries automatically. Legitimate repeated code/text may trigger the heuristic.
Time limits are cooperative between generation steps, not a hard process deadline.
The helper is opt-in: plain Transformers, GGUF and Ollama imports do not automatically
execute it. These safeguards do not change weights and do not fix factual errors.
Selected regression tests are documented in RUNTIME-PATCH-VALIDATION.json; they
are not independent benchmark scores or evidence of universal improvement.
📁 Files to use
Standard model-*.safetensors, model-mtp.safetensors and the index load together with tokenizer, processor, chat template and configs. Keep standard filenames unchanged. About 70 GB of weights plus runtime overhead; use a quantized edition for constrained memory. q36_runtime.py is optional and batch-size-one only, not an automatic GGUF/Ollama patch.
🧰 Support files and known limitations
SHA256SUMS verifies release files. RELEASE-NAMING.json maps old repository/file names to v2 and records unchanged weight hashes. Historical validation reports intentionally retain their original runtime names; the map resolves those names. RELEASE-VALIDATION.json covers tensor/merge/provenance checks; RUNTIME-PATCH-VALIDATION.json covers selected guard tests.
Thinking mode can loop; start with thinking off and bounded output. Guard stops are incomplete answers, not successful corrections. The private training inputs, installer ZIP, raw benchmark prompts/answers and internal debug logs are not part of these end-user repositories. Keep use within applicable rights and deployment requirements.
📊 Measured diagnostics — scope matters
These are local Q4_K_M / llama.cpp diagnostics, not full official benchmark scores and not Ollama performance claims. Temperature 0, seed 42, thinking off, 16,384 context, 4,096 output cap, 45-second per-request budget. The RTX 2000 Ada used requested 24 GPU layers plus CPU offload; the actual offload-layer log was unavailable.
Table with columns: Model / device, Attempted / 204, Completed, Old strict pass / scored, End-to-end output tok/s (median, outputs ≥64 tokens)| Model / device | Attempted / 204 | Completed | Old strict pass / scored | End-to-end output tok/s (median, outputs ≥64 tokens) |
|---|
| Q36 / H200 | 204 | 186 | 61 / 108 | 156.4 |
| Q36 / RTX 5090 | 204 | 190 | 62 / 109 | 174.6 |
| Huihui / RTX 5090 | 204 | 196 | 87 / 114 | 203.8 |
These strict counts omit pending official/manual evaluators, use changing denominators, and sometimes reject semantically correct formatting variants. Do not divide passes by all prompts or call these counts overall accuracy. A separate local regrade keeps content, protocol compliance and delivery apart. It does not certify unreviewed reasoning or execute generated code. The Huihui RTX 2000 Ada run left 22 tasks untested at the global deadline. Earlier BF16 diagnostics used thinking and are not directly comparable.
Huihui performed better on the shared automatically assessable RTX 5090 subset; no claim is made that this fine-tune universally surpasses its source or Qwen3.8. Rates mix generated lengths and are not pure hardware speedups. Native decode, prefill, client first-output and resource metrics are separated in BENCHMARK-DEVICE-SUMMARY.json. Raw prompts/answers and private training data are not published here.
Qwen3.8 comparison plan registers every benchmark family from the publisher card, including internal/unavailable tasks. Qwen3.8 has not been tested locally. Publisher scores use different harnesses, settings, annotations and trial counts and are shown only as references. No GPU job is launched by these support files.
🔎 Broad benchmark — tests and passes
Expand a device to inspect every test family. Passes / graded use the original strict evaluator, including format-sensitive checks. Ungraded answers are not failures or passes. Incomplete is a separate delivery flag and can overlap with ungraded. These are small local subsets, not official leaderboard scores.
Table with columns: Test family, Attempted / planned, Passes / graded, Ungraded, Incomplete, Not run| Test family | Attempted / planned | Passes / graded | Ungraded | Incomplete | Not run |
|---|
| ARC-Challenge | 10 / 10 | 8 / 8 | 2 | 2 | 0 |
| BBH | 23 / 23 | 3 / 20 | 3 | 3 | 0 |
| ChartQA | 5 / 5 |
Table with columns: Test family, Attempted / planned, Passes / graded, Ungraded, Incomplete, Not run| Test family | Attempted / planned | Passes / graded | Ungraded | Incomplete | Not run |
|---|
| ARC-Challenge | 10 / 10 | 8 / 8 | 2 | 2 | 0 |
| BBH | 23 / 23 | 3 / 20 | 3 | 3 | 0 |
| ChartQA | 5 / 5 |
Table with columns: Test family, Attempted / planned, Passes / graded, Ungraded, Incomplete, Not run| Test family | Attempted / planned | Passes / graded | Ungraded | Incomplete | Not run |
|---|
| ARC-Challenge | 10 / 10 | 10 / 10 | 0 | 0 | 0 |
| BBH | 23 / 23 | 16 / 23 | 0 | 0 | 0 |
| ChartQA | 5 / 5 |
Table with columns: Test family, Attempted / planned, Passes / graded, Ungraded, Incomplete, Not run| Test family | Attempted / planned | Passes / graded | Ungraded | Incomplete | Not run |
|---|
| ARC-Challenge | 10 / 10 | 7 / 7 | 3 | 3 | 0 |
| BBH | 23 / 23 | 3 / 20 | 3 | 3 | 0 |
| ChartQA | 5 / 5 |
Table with columns: Test family, Attempted / planned, Passes / graded, Ungraded, Incomplete, Not run| Test family | Attempted / planned | Passes / graded | Ungraded | Incomplete | Not run |
|---|
| ARC-Challenge | 10 / 10 | 10 / 10 | 0 | 0 | 0 |
| BBH | 11 / 23 | 6 / 10 | 1 | 1 | 12 |
| ChartQA | 5 / 5 |
Machine-readable counts. Code generation is not a pass until the relevant execution tests have been graded.
🏁 Hard benchmarks on the roadmap
All 25 families below are registered from the Qwen3.8-27B card. The matching official adapters and datasets are not yet fully prepared, and no local Qwen3.8 baseline has been run. Existing short similarly named diagnostics do not substitute for those runs.
Table with columns: Benchmark, Planned local cases, Preparation status| Benchmark | Planned local cases | Preparation status |
|---|
| Terminal Bench 2.1 (Terminus) | 5 | Adapter/data preparation pending |
| SWE-bench Pro | 5 | Adapter/data preparation pending |
| NL2Repo-Bench | 5 | Adapter/data preparation pending |
| DeepSWE 1.1 | 5 | Adapter/data preparation pending |
| QwenSWEBench | TBD | Internal release/access needed |
| CoWorkBench |
Comparison graphics will follow measured results, with our local samples and publisher-reported scores kept clearly separate. Different harnesses, budgets and trial counts will not be presented as a head-to-head win. Full protocol and references.
💻 Coding, agents and your laptop
The next local checks cover executable coding tests, bug fixes, tool calls, planning, recovery, memory and stopping at the right time. 64 case slots are specified, including 19 existing coding slots. They are compact skill diagnostics, not proof of AGI. Results and failures will both be reported; new capability claims require actual task-level evidence.
🛠️ Laptop test plan, Hermes routing and Obsidian workflows. The runtime comparison and full memory integration are in preparation, not yet validated. The four device quickstarts above remain separate from these future end-to-end checks.
📜 License, lineage & credits
Apache-2.0 license. Dataset sources and their own license information are linked in DATASETS.md; model licensing does not relicense the source datasets. Thanks to the Qwen team, the inherited model and dataset authors, and the Soup, PEFT, Transformers, llama.cpp and Ollama projects.
Name guide: Qwen3.6-35B identifies the model family and approximate total parameter count; v2 is this project release; MoE means mixture of experts; Ablit, Heretic, Uncensor and Hermes describe inherited project lineage/training intent. MTP denotes preserved multi-token prediction weights, not a measured speedup; Vision denotes image-input support. FT, Llama and Ollama distinguish the three packages.
These names do not imply affiliation or a promise of unrestricted behavior. Full provenance and validation are retained, including the historical Opus4.7-labelled source. No universal superiority, guaranteed compliance or removal of memorization is claimed.
This release builds on the previous project's model card, which records the following stages. These are inherited stages, not new operations performed while packaging this release.
Table with columns: Name / stage, What it contributes| Name / stage | What it contributes |
|---|
| Qwen3.6 / MoE | The underlying language-and-vision architecture and mixture-of-experts backbone. |
| Reasoning-distilled lineage | The earlier card traces a lordx64 derivative followed by the huihui-ai derivative. The historical Opus4.7 label describes inherited reasoning-distillation provenance, not inclusion of proprietary Claude weights. |
| Ablit / abliterated | The huihui-ai source underwent a weight-modification stage aimed at reducing refusal behavior. It is not a separate tool library. |
| Heretic | A subsequent custom, fused-MoE-aware modification stage documented by the previous release. It is distinct from the earlier abliteration. |
| OBLITERATUS Nuclear / Uncensor | Another separately recorded inherited modification stage aimed at reducing refusal behavior. Uncensor is the concise release-name label; OBLITERATUS remains part of the provenance. |
| Hermes | Tool-oriented supervised training: function-call structure, coding, terminal/file/repository workflows and multi-tool coordination, using and . Your host application still supplies, authorizes and executes tools. |
The old card reports 23,220 training and 1,179 validation examples for its own earlier SFT stage; those are separate from the current release's 34,000 record uses. It also records protection of 333 vision tensors and 19 MTP tensors in that earlier build. These historical checks are not new laptop benchmark results.
The previous release remains separate: Qwen3.6 Opus4.7 Heretic Hermes Agent — Editions.