---

base_model:

• flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed

base_model_relation: quantized

library_name: gguf

license: apache-2.0

license_link: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/LICENSE

pipeline_tag: text-generation

language:

• en
• ru

tags:

• gguf
• llama.cpp
• qwen3.6
• qwen3_5_moe
• reasoning
• tool-calling
• agentic
• coding
• bilingual
• imatrix

datasets:

• open-r1/OpenR1-Math-220k
• simplescaling/s1K-1.1
• NousResearch/hermes-function-calling-v1
• glaiveai/glaive-function-calling-v2
• OpenAssistant/oasst2

quantized_by: flashback2k

---

⚡ FlashModel-Qwen-3.6-35B-A3B-Maxed — GGUF

GGUF quants of FlashModel-Qwen-3.6-35B-A3B-Maxed,

a reasoning / coding / agentic / bilingual fine-tune of Qwen/Qwen3.6-35B-A3B.

Key capabilities:

• Code: Rust, TypeScript, JavaScript, Python, QML, HTML, CSS, Bash
• Controllable reasoning effort: reasoning_effort=low|medium|high|xhigh|max in the system prompt — real control, not truncation
• Agentic tool use: file read/write, shell, git operations
• Bilingual EN/RU dialogue (≈22% RU in training)
• MTP speculative decoding (--spec-type draft-mtp): 77% draft acceptance; measured decode 29.2 vs 27.2 tok/s without MTP (about +7%, one 400-token sample, IQ3_XXS on a Radeon 680M iGPU).

Provided Quants

All quants use an importance matrix computed on this fine-tune with code-weighted calibration (local coding chats + Rust/TS/JS/Python/QML/CSS samples, plus some RU).

Every tensor uses the quant's base type (stock llama-quantize, no expert overrides); the MTP head is kept at Q4_K/Q5_K in the low-bit quants. Sizes include the MTP head.

Why this range? Below IQ3_XXS quality degrades noticeably on code and reasoning; above

Q5_K_M the gain is too small to justify the extra VRAM — for a 24 GB card the sweet spot is

already covered. Q4_K_M is the default recommendation.

| Link | Type | Size/GB | Notes |

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

| GGUF | IQ3_XXS | ~14.1 | floor — smaller starts to hurt quality |

| GGUF | IQ3_S | ~15.4 | better quality |

| GGUF | IQ4_XS | ~18.2 | good trade-off |

| GGUF | IQ4_NL | ~18.5 | balanced+ |

| GGUF | Q4_K_M | ~22.7 | recommended for 24 GB VRAM |

| GGUF | Q5_K_S | ~24.9 | 24 GB VRAM + offload some layers |

| GGUF | Q5_K_M | ~26.5 | ceiling — higher is pointless for this VRAM |

| imatrix | imatrix | — | importance matrix used for every quant above |

Which one should I pick?

| Your free VRAM | Pick |

|:--|:--|

| ≥ 28 GB | Q5_K_M — best quality available |

| ~26 GB | Q5_K_S |

| 24 GB | Q4_K_M — default choice |

| ~22 GB | IQ4_XS |

| ≤ 20 GB | IQ3_S, then IQ3_XXS |

Leave room for the KV cache: long reasoning traces need context. Size numbers are file sizes, not total memory use.

Context length beyond 16k is not yet measured on this fine-tune.

mmproj

mmproj-Q8_0.gguf — vision encoder in Q8_0 (~0.45 GB). Use with vision-enabled builds.

Usage

llama.cpp (recommended)

# download one quant
huggingface-cli download flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed-GGUF \
  FlashModel-Qwen-3.6-35B-A3B-Maxed-Q4_K_M.gguf mmproj-Q8_0.gguf --local-dir .

# server with MTP speculative decoding
llama-server -m FlashModel-Qwen-3.6-35B-A3B-Maxed-Q4_K_M.gguf \
  --mmproj mmproj-Q8_0.gguf --spec-type draft-mtp --spec-draft-n-max 2 \
  -c 32768 --flash-attn -ngl 99 --jinja

# effort control via system prompt
#   low    -> 1-2 sentences
#   medium -> paragraph
#   high   -> multi-paragraph reasoning
#   max    -> full chain-of-thought
system: "You are a coding and reasoning assistant.
reasoning_effort=high
Adjust the depth of your reasoning to the level given above."

Or pull straight from the Hub: llama-server -hf flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed-GGUF:Q4_K_M --jinja

Ollama

ollama run hf.co/flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed-GGUF:Q4_K_M

LM Studio

Search for flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed-GGUF and pick a quant.

Recommended sampling (from Qwen3.6)

| Mode | temperature | top_p | top_k | min_p | presence_penalty |

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

| Thinking, general | 1.0 | 0.95 | 20 | 0.0 | 1.5 |

| Thinking, precise coding | 0.6 | 0.95 | 20 | 0.0 | 0.0 |

| Non-thinking, general | 0.7 | 0.8 | 20 | 0.0 | 1.5 |

Always pass --jinja so the embedded chat template (thinking blocks, tool calls) is used.

How these were made

Base: flashback2k/FlashModel-Qwen-3.6-35B-A3B-Maxed

(HF bf16 safetensors, LoRA merged, MTP head kept from the Qwen3.6-35B-A3B checkpoint).

Converted with llama.cpp/convert_hf_to_gguf.py to F16 (NextN/MTP tensors included), then

quantized with stock llama-quantize using an imatrix computed on this fine-tune (code-weighted calibration).

No per-tensor expert overrides; the MTP head is kept at Q4_K/Q5_K in the IQ3 quants.

No per-tensor expert overrides: each quant uses its base type for all tensors. The MTP head (block 40) is kept at Q4_K/Q5_K in the IQ3 quants because it never runs in a normal forward pass and so has no imatrix data.

Known eval results (baseline → finetuned)

| Metric | Baseline | Finetuned |

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

| Tool Use (8 scenarios) | 6/8 | 7/8 |

| Agentic (5 tasks) | 5/5 | 5/5 |

| Effort monotonic (low→max) | ❌ | ✅ |

| RU Perplexity | — | 5.41 |

Full benchmark suite (HumanEval, MBPP, GSM8K, MMLU, SWE-bench) planned — see the weights repo card for training details.