DEV.co
Open-Source LLM · hesamation

Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF

A 35B parameter reasoning-focused LLM fine-tuned via supervised learning on Claude Opus chain-of-thought data. Distributed as GGUF quantizations for local inference via llama.cpp. Apache 2.0 licensed, non-gated. Shows +32.85 percentage-point improvement on MMLU-Pro versus base model.

Source: HuggingFace — huggingface.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
40.6k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerhesamation
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads40.6k
Likes285
Last updated2026-04-19
Sourcehesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF

What Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF is

Qwen3.6-35B-A3B base model fine-tuned using LoRA (rank 32, attention-only) on ~14,200 reasoning-distillation samples from three datasets (Claude Opus trajectories, curated Qwen examples, additional Claude reasoning). Trained for 2 epochs over 762 steps with 8192-token max sequence length and 32768 context window support. Merged and quantized to GGUF format (Q4_K_M, Q5_K_M, Q6_K, Q8_0 variants) for llama.cpp inference. Text-only fine-tuning; vision encoder inherited from base but not tuned.

Quickstart

Run Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Local reasoning-heavy workloads

Fine-tuning demonstrates +32.85pp gain on MMLU-Pro. Suitable for on-premise QA, code reasoning, or math-heavy applications where chain-of-thought helps accuracy and no cloud inference is acceptable.

Cost-sensitive deployment with VRAM constraints

Multiple GGUF quantizations (Q4–Q8) allow deployment on consumer/edge hardware. Q4_K_M targets smallest practical local inference; Q5–Q6 offer better quality/size trade-off.

Custom fine-tuning starting point

Reasoning-distilled checkpoint can serve as base for further LoRA tuning on domain-specific reasoning tasks (legal analysis, scientific problem-solving). Unsloth-compatible training stack noted in acknowledgements.

Running & fine-tuning it

ESTIMATE ONLY—verify for your setup: Q4_K_M ~14–16 GB VRAM (consumer GPU; low-end server); Q5_K_M ~18–20 GB; Q6_K ~22–24 GB; Q8_0 ~28–32 GB (closest to fp16 baseline). Base model 35B fp16 would require ~70 GB. Exact footprint depends on llama.cpp batch size, context length, and quantization implementation. CPU inference possible but slow.

Source model trained with LoRA (rank 32, alpha 32, attention-only) using Unsloth. Merged checkpoint available. Further LoRA tuning on merged model is feasible; re-quantizing after tuning requires llama.cpp toolchain. Low LoRA rank (32) suggests conservative budget; may be suitable for downstream LoRA on resource-limited setups. QLoRA compatibility unknown; not stated in card.

When to avoid it — and what to weigh

  • Multimodal/vision requirements — Base Qwen3.6 includes vision encoder but this fine-tune is text-only. Image/video handling not improved by training; treat as base model capabilities only.
  • Inference with latency-critical SLAs — 35B parameter model requires substantial compute. Q4_K_M quantization reduces quality; Q6–Q8 demand significant VRAM. Production latency depends heavily on quantization and hardware.
  • Minimal documentation or training details needed — Card provides training hyperparameters and dataset sources but limited evaluation beyond one smoke-test MMLU-Pro run (70 samples with --limit 5). Quantization impact on downstream performance not characterized.
  • Long-running production systems without community validation — Model is community-maintained by single developer. No official benchmarks on real-world tasks. Card explicitly welcomes independent community benchmarks; limited production reference deployments likely.

License & commercial use

Apache 2.0 (OSI-approved permissive license). No usage restrictions on commercial or proprietary application.

Apache 2.0 is an OSI-permissive license allowing commercial use, modification, and distribution with attribution and liability disclaimers. No commercial-use restrictions. However: (1) Base model (Qwen3.6-35B-A3B) and training data (Claude Opus, third-party reasoning datasets) may have separate license/TOS. Verify Qwen team and dataset owner licenses independently. (2) Community-maintained model; no SLA or commercial support offered. (3) Derivative works (GGUF quantizations) inherit Apache 2.0 but ensure upstream base-model license chain is compatible for your use case.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceUnknown
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceMedium
Security considerations

No explicit security analysis provided in card. Considerations: (1) Training data sourced from Claude Opus reasoning and community-curated datasets; potential for biased or harmful outputs inherited from training data. (2) Model card does not address jailbreak resilience, adversarial input handling, or content-filtering. (3) Local inference reduces data exfiltration risk vs. cloud APIs but operator is responsible for access control and model safety. (4) GGUF quantization is mechanically deterministic (no novel attack surface) but should be run in isolated/sandboxed environments if used on untrusted input.

Alternatives to consider

Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled (Jackrong)

Smaller (27B vs. 35B), same reasoning-distillation approach, likely lower VRAM footprint. Fewer downloads (40k for this model) suggests Jackrong version may have earlier validation; trade-off is parameter count.

Llama 3.1 (Meta) with LoRA fine-tuning on reasoning data

Larger official community (Meta-backed), extensive tooling (vLLM, TGI native support), clearer licensing. Requires own LoRA tuning if reasoning-specific behavior desired; more development overhead but better production readiness.

Mixtral 8x22B or similar MoE models

Base Qwen3.6 includes MoE (qwen3_moe tag noted). Mixtral offers comparable reasoning capability with sparser inference (faster latency) and wider ecosystem support. Larger context, more mature quantization ecosystem.

Software development agency

Ship Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF with senior software developers

Evaluate this model for cost-sensitive, on-premise reasoning workloads. Start with Q5_K_M quantization (~18–20 GB VRAM) and benchmark on your task before production rollout. Community-maintained; Devco can assist with integration, fine-tuning, and deployment validation.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF FAQ

Can I use this commercially?
Apache 2.0 permits commercial use with attribution. However, verify upstream licenses: Qwen/Qwen3.6-35B-A3B and the three training datasets (Claude Opus reasoning data, etc.) must also be compatible for your use case. Some Claude Opus data may be proprietary; review Anthropic TOS and dataset owners' licenses.
What GGUF quantization should I choose?
Start with Q4_K_M if VRAM is <16 GB (smallest/fastest, quality trade-off). Q5_K_M (18–20 GB) offers better quality/size balance for most local inference. Q6_K or Q8_0 if you have >24 GB VRAM and prioritize quality. Card notes this is source-checkpoint guidance; quantization impact on downstream tasks not independently benchmarked.
How do I deploy this with llama.cpp?
Download a GGUF file, then run: `./main -m <model.gguf> -n 512 -p "Your prompt"`. Qwen3.6 architecture requires recent llama.cpp build (check releases for qwen3 support). For production, use llama-cpp-python or other bindings. Context length up to 32k supported by base model; adjust -n and -c flags as needed.
What does the +32.85pp MMLU-Pro improvement mean?
Base model scored 42.86% on a 70-question smoke-test (--limit 5); fine-tuned model scored 75.71%. This is not a full MMLU-Pro benchmark (which is 14k questions) and used aggressive limiting. Treat as comparative check, not absolute accuracy guarantee. Independent evaluation strongly recommended before production use.

Software development & web development with DEV.co

DEV.co helps companies turn open-source tools like Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Deploy Reasoning-Enhanced LLM Locally

Evaluate this model for cost-sensitive, on-premise reasoning workloads. Start with Q5_K_M quantization (~18–20 GB VRAM) and benchmark on your task before production rollout. Community-maintained; Devco can assist with integration, fine-tuning, and deployment validation.