--- base_model: Qwen/Qwen3.6-35B-A3B base_model_relation: quantized quantized_by: IsValorum library_name: gguf pipeline_tag: image-text-to-text license: apache-2.0 language: - en - zh - es - fr - de - pt - it - ru - ja - ko - vi - th - ar tags: - gguf - llama.cpp - quantized - quantization - apex - apex-quant - apex-i-nanoplus - nanoplus - custom-quantization - unsloth-studio - qwen3.6 - qwen35moe - qwen3_5_moe - qwen - moe - reasoning - conversational - multimodal - vision - image-text-to-text - coding - agentic - swe-bench - mtp - speculative-decoding --- ## Quick Navigation Index 1. [Optimization History & Transparency Notice](#toc-01) 2. [Independent Benchmark of the APEX-I-MiniPlus Family (Occamy V2 Reference)](#toc-02) 3. [Empirical Benchmarks & Fidelity Verification](#toc-03) 4. [Quality Spectrum: APEX-I-NanoPlus vs. Standard Flat Quantizations](#toc-04) 5. [Model Files & Technical Specifications](#toc-05) 6. [Surgical Tensor Quantization Map (Audited from GGUF)](#toc-06) 7. [Hardware Throughput & Offload Benchmarks (RTX 30 / 40 / 50 & RAM Streaming)](#toc-07) 8. [The 16GB & 24GB Advantage: Massive Context in VRAM!](#toc-08) 9. [Recommended Configuration & Setup](#toc-09) 10. [1. llama-cli with MTP Speculative Decoding & Vision Projector](#toc-10) 11. [2. High-Performance OpenAI-Compatible Server](#toc-11) 12. [Recommended Generation Parameters (Qwen Official)](#toc-12) 13. [CRITICAL: Coding Syntax & Repeat Penalty Advisory (Preventing Character Swapping)](#toc-coding-advisory) 14. [Hardened Agentic Chat Template & Reasoning Effort](#toc-chat-template) 15. [Optional Support](#toc-13) # Qwen3.6-35B-A3B-MTP APEX-I-NanoPlus GGUF ### *The Next-Generation Frontier MoE Β· Extreme 13GB Footprint Β· Fast System RAM Streaming & Massive Context on 16GB VRAM* > [!IMPORTANT] > ### THE DEFINITIVE SPECIFICATION IN THE approx. 13 GB CEILING > This **APEX-I-NanoPlus** release marks the official debut of our specialized tensor-by-tensor architectural configuration for sparse Mixture-of-Experts quantization within an extreme **approx. 13 GB envelope**. Its 40 layers and 256 micro-experts are allocated tensor by tensor to maximize reasoning precision, preserve routing behavior, and prevent avoidable CPU dequantization stalls during hybrid and system-RAM inference; optional MTP companion files remain available for speculative decoding. > [!TIP] > ### πŸ† EMPIRICAL BENCHMARK & QUALITY COMPARISON > > | Quantization Specification | File Size (Disk) | Memory Footprint (RAM/VRAM) | Average BPW | WikiText-2 Perplexity | Quality Tier Equivalent | > | :--- | :---: | :---: | :---: | :---: | :---: | > | **Unquantized BF16 Base** | approx. 71.05 GB | approx. 66.18 GiB | 16.00 BPW | approx. 5.32 | Full Precision Baseline | > | **[APEX-I-MiniPlus V2.1](https://huggingface.co/IsValorum/Qwen3.6-35B-A3B-MTP-APEX-I-MiniPlus-V2.1-GGUF)** | 15.23 GB | 14.18 GiB | 3.43 BPW | 5.3693 Β± 0.12528 (+0.0493 / +0.93%) | Q5_K_L tier (bordering Q6_K) | > | **APEX-I-NanoPlus (NEW)** | **13.03 GB** | **12.14 GiB** | **2.93 BPW** | **5.5244 Β± 0.12916 (+0.2044 / +3.84%)** | **Solid Q4_K_M / Q4_K_L Tier** | > | *Generic Community IQ2_S* | approx. 12.2 GB | approx. 11.4 GiB | 2.56 BPW | > 5.85 (Degraded) | Unstable / Syntax Spikes | > > **Looking for higher precision?** [APEX-I-MiniPlus V2.1](https://huggingface.co/IsValorum/Qwen3.6-35B-A3B-MTP-APEX-I-MiniPlus-V2.1-GGUF) offers the full **15.23 GB (3.43 BPW)** release of this Qwen3.6 family, delivering Q5_K_L tier (bordering Q6_K) fidelity with integrated MTP support. > > **ARC-Challenge (0-shot, 1,172 questions): approx. 95.75%.** > - **Q4_K_L Tier in Reasoning & Routing:** 100% uncompressed `F32` routers (`gate_inp`) and a `Q6_K` output head eliminate router drift, matching or exceeding standard `Q4_K_L` baselines on logic benchmarks. > - **Solid Q4_K_M Tier in Language Modeling:** WikiText-2 perplexity preserves 4-bit distributional fidelity across standard generation in an agile approx. 13.03 GB footprint. > > **Routing:** all recipe-designated `gate_inp` and `gate_shexp` tensors remain in uncompressed `F32`, preserving zero routing drift. > [!WARNING] > ### DO NOT CONFUSE APEX-I-NANOPLUS WITH GENERIC COMMUNITY SUB-3-BIT QUANTS! > **Regardless of release version, NEVER confuse handcrafted APEX-I-NanoPlus builds with generic community sub-3-bit releases:** > - **Generic Community IQ2_S / IQ2_XXS:** Uniformly crushes all core MoE experts down to aggressive 2-bit codebooks without importance calibration, leaves the sensitive token output head unarmored at 3-bit, and compresses attention projections. In deep reasoning models, this triggers severe perplexity spikes, syntax errors, and broken code brackets. > - **Handcrafted APEX-I-NanoPlus:** Applies a surgical tensor-by-tensor architecture that preserves 100% of expert routing matrices in uncompressed `F32` (zero router drift), armors the token output head in high-precision `Q6_K`, safeguards attention gates in `Q8_0`, fortifies the critical MoE down-projection residual stream (`ffn_down_exps`) in `IQ3_XXS` (3.06 bpw), and restricts 2-bit compression strictly to redundant gating/up projections guided by the official `imatrix`. > [!TIP] > ### SYSTEM RAM INFERENCE: FULL OR PARTIAL > This APEX-I-NanoPlus release is designed for **full or partial system-RAM inference**. Depending on the processor, memory bandwidth, and DDR4/DDR5 configuration, generation can range from **20 to 45 tok/s**. With partial GPU offload, systems that cannot fit **128K or more context** entirely in VRAM can place the remaining model and context load in system RAM, maintaining stable, responsive generation at longer context lengths. --- ## Optimization History & Transparency Notice We maintain our previous releases publicly as a transparent engineering record of continuous optimization. Below is the exact evolutionary roadmap of our architectures: | Specification | Core Experts (2–37) | Edge Experts (0–1, 38–39) | Shared Expert (`shexp`) | Full Attention (L3, 7, 11, ...) | Attention Gates (30 Layers) | Output Head (`output.weight`) | Routers (`gate_inp`) | Size / Overhead | Real-World Impact | | :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :--- | | **Generic APEX Mini** | `IQ2_S` (2.50 bpw) | `Q3_K` (only 5 layers) | `Q4_K` / `Q3_K` | `Q3_K` | Compressed | `Q3_K_M` | Compressed | Baseline (approx. 12.5 GB) | Severe syntax errors, broken code indentation, high perplexity in ``. | | **MiniPlus V2.1 (Current)** | `IQ3_XXS` + `Q3_K` | `Q3_K` (10 layers) | `Q5_K` | `Q4_K` (`q/k/v`) + `Q6_K` (`output`) | `Q8_0` | `Q6_K` | `F32` | 15.23 GB (14.18 GiB) | Maximum fidelity near-lossless Q5/Q6 tier. Fits 24GB GPUs effortlessly. | | **NanoPlus (NEW)** | **`IQ3_XXS` (down) + `IQ2_S` (gate) + `IQ2_XXS` (up)** | **`Q3_K` (down) + `IQ3_XXS` (gate/up)** | **`Q4_K`** | **`Q4_K` (`q/k/v`) + `Q6_K` (`output`)** | **`Q8_0`** | **`Q6_K`** | **`F32`** | **13.03 GB (12.14 GiB)** | **Calibrated approx. 13.0 GB tier. Leaves >3 GB free VRAM on 16GB cards for 32k context with zero AVX2 CPU stalls.** | > [!TIP] > ### Deployment & System Architecture Guide > - **Full GPU VRAM Offload (16GB+ VRAM, `-ngl 99`):** Effortless full offload with native 32K–64K context support on 16GB cards (RTX 4080 / RTX 4070 Ti Super), and native 256K context on 24GB workstations (RTX 3090 / 4090 / 5090). > - **System RAM Streaming Specialist (DDR4/DDR5 & Massive Context):** Specially engineered to run either partially or entirely out of system RAM across large or full context windows. By utilizing linear SIMD-optimized `Q4_K` attention projections and preserving critical down-projections in `IQ3_XXS`, AVX2 CPU dequantization stalls are eliminated. Depending on your processor architecture and memory bandwidth (dual-channel DDR4 or high-speed DDR5 6000+ MT/s), streaming generation in system RAM can approach speeds remarkably close to full VRAM execution. > > Explore our official collection: > **[APEX-I-NanoPlus Collection](https://huggingface.co/collections/IsValorum/apex-i-nanoplus-6ab41467c988a1b1cb9b83bc)**. --- > [!IMPORTANT] > ### EXPLORE THE ESTABLISHED 35B MoE MINIPLUS LINEUP > These are complementary APEX-I-MiniPlus V2.1 releases, not alternate downloads of the same model. Each receives the same tensor-by-tensor approach, integrated MTP where supported, and a design suitable for full or partial system-RAM inference. Choose the model whose native strengths best fit the work you want to do: > > - **[Qwen3.6-35B-A3B MTP APEX-I-MiniPlus-V2.1](https://huggingface.co/IsValorum/Qwen3.6-35B-A3B-MTP-APEX-I-MiniPlus-V2.1-GGUF)** β€” a versatile frontier MoE for broad reasoning, multilingual work, agents, tool use, and multimodal tasks. > - **Best for:** General reasoning, agent workflows, tool calling, and flexible multimodal use. > - **[Qwen3.6-35B-A3B MTP APEX-I-MiniPlus-V2.1 Abliterated](https://huggingface.co/IsValorum/Qwen3.6-35B-A3B-MTP-APEX-I-MiniPlus-V2.1-Abliterated-GGUF)** β€” the V2.1 refusal-ablated Qwen3.6 edition for users who deliberately prefer reduced refusal behavior. > - **Best for:** Workflows where an abliterated Qwen3.6 variant is explicitly desired. > - **[Ornith 1.5 APEX-I-MiniPlus-V2.1](https://huggingface.co/IsValorum/Ornith-1.5-35B-A3B-APEX-I-MiniPlus-V2.1-GGUF)** β€” a software-engineering-focused MoE designed for repository-scale coding and autonomous engineering agents. > - **Best for:** Repository-scale development, multi-file code changes, and software-engineering agents. > - **[Tiel Coder APEX-I-MiniPlus-V2.1](https://huggingface.co/IsValorum/Tiel-Coder-35B-A3B-APEX-I-MiniPlus-V2.1-GGUF)** β€” a specialist coding MoE tuned for agentic programming, iterative tool use, and implementation-heavy work. > - **Best for:** Focused coding sessions, iterative debugging, and tool-driven implementation. > > These remain distinct model families and editions with their own behavior and empirical results. Pick by workload and intended alignment behavior rather than treating them as interchangeable quantization variants. --- ### πŸ… Independent Benchmark of the APEX-I-MiniPlus Family (Occamy V2 Reference) > [!NOTE] > **External report:** [zephel01 independently benchmarked Occamy V2](https://note.com/zephel01/n/n71d3d7e6b70c?hl=en). The benchmark was performed on **Occamy-1.0 APEX-I-MiniPlus V2**, not on this specific NanoPlus release. It is included as independent context for the broader tensor-by-tensor MiniPlus approach. The related **APEX-I-MiniPlus** quantization architecture has an independent evaluation by Japanese AI researcher and evaluator [zephel01 (CoolZero)](https://note.com/zephel01/n/n71d3d7e6b70c?hl=en) on an **NVIDIA RTX 5090 (32GB)** workstation running `llama.cpp` CUDA with FlashAttention. The evaluation covered **348 unseeded trials** on SWE-bench-style multi-file Python bug-fixing tasks with hidden `pytest` suites (`llmbench`): - **L6 Multi-File Code Generation (60 tasks):** - **32K context:** **93.3% resolved** (46/60 tasks passed 5/5 consecutive trials). - **65K context:** **90.0% resolved** (45/60 tasks passed 5/5 consecutive trials). - **Context scaling on the tested Occamy V2 build:** 14.6 GB at 32K and 15.1 GB at 65K. These findings belong to Occamy V2 and should not be read as a direct NanoPlus benchmark. NanoPlus-specific fidelity is documented by its own WikiText-2 result below. ## Empirical Benchmarks & Fidelity Verification The comparison table near the top consolidates the model-specific BF16 baseline, final GGUF PPL, delta, published main-file size, BPW, and fidelity tier. The routing treatment is preserved in the note directly beneath it. ### Quality Spectrum: APEX-I-NanoPlus vs. Standard Flat Quantizations How the handcrafted **APEX-I-NanoPlus** architecture compares against standard flat quantizations in `llama.cpp` on 35B Mixture-of-Experts architectures: | Quantization Format | Bits Per Weight (BPW) | Model Footprint (Disk / VRAM) | Perplexity Delta (vs. FP16 Baseline) | Token Fidelity & Syntactic Stability Tier | | :--- | :---: | :---: | :---: | :--- | | **FP16 / BF16 (Uncompressed)** | 16.0 bpw | 71.0 GB | **0.00** (Reference) | 100% full uncompressed reference fidelity. | | **Standard Q8_0** | 8.50 bpw | approx. 38 GB | approx. +0.01 | Virtually lossless; excessive memory overhead for consumer hardware. | | **Standard Q6_K** | 6.56 bpw | approx. 30 GB | approx. +0.02 to +0.05 | Near-lossless FP16 fidelity; requires multi-GPU or 32GB+ VRAM setups. | | **APEX-I-MiniPlus V2.1** | 3.43 bpw | 15.23 GB (14.18 GiB) | **+0.0493 (PPL: 5.3693)** | Near-lossless Q5_K / Q6_K tier. Full native 256K context on 24GB workstations. | | πŸ† **APEX-I-NanoPlus (IsValorum)** | **2.93 bpw** | **13.03 GB (12.14 GiB)** | **+0.2044 (PPL: 5.5244 Β± 0.12916)** | **Solid Q4_K_M fidelity tier at only 13.0 GB (81.6% weight reduction). Enables 32k context 100% in VRAM on 16GB GPUs.** | | **Standard Q4_K_M** | 4.50 bpw | approx. 20.0 GB | approx. +0.15 to +0.25 | Standard industry trade-off; cannot fit in 16GB VRAM. | | **Standard Q3_K_M** | 3.44 bpw | 16.6 GB | approx. +0.36 to +0.45 | Noticeable syntax drop, bracket corruption, and tokenizer classification noise. | | **Standard IQ2_S / Generic APEX Mini** | 2.50 bpw | approx. 12.2 GB | approx. +0.55 to +1.50+ | Severe reasoning breakdown, high perplexity spikes in `` chains. | --- ## Model Files & Technical Specifications | File Name | File Size | Memory Footprint | BPW | Description | | :--- | :--- | :--- | :--- | :--- | | **`Qwen3.6-35B-A3B.APEX-I-NanoPlus.gguf`** | **`13.03 GB (12.14 GiB)`** | `12.14 GiB` | **2.93 BPW** | Core agentic reasoning & multimodal MoE in APEX-I-NanoPlus | | **`mmproj-Q8_0.gguf`** | **`610.66 MB (582.37 MiB)`** | `582.37 MiB` | **8.50 BPW** | Dedicated Q8_0 multimodal vision projector for document & image reasoning | | **`mtp-Qwen3.6-35B-A3B-Q8_0.gguf`** | **`1.99 GB (1.85 GiB)`** | `1.85 GiB` | **8.50 BPW** | Dedicated Q8_0 Multi-Token Prediction (MTP) draft head for speculative decoding | | **`mtp-Qwen3.6-35B-A3B-Q4_0.gguf`** | **`1.19 GB (1.11 GiB)`** | `1.11 GiB` | **4.50 BPW** | Dedicated Q4_0 Multi-Token Prediction (MTP) draft head for low-VRAM speculative decoding | | **Complete download** | **`14.83 GB (13.81 GiB)`** | `13.81 GiB` | β€” | Main GGUF plus the Q4_0 MTP companion and bundled vision projector | - **Base Model:** [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) - **Parameters:** 35.2B total (approx. 2.6B to 3.2B active per token) - **Architecture:** 40 layers, 256 micro-experts (8 active per token) + hybrid linear attention / DeltaNet recurrent layers; separate MTP companion heads are available for speculative decoding - **Context Length:** 262,144 tokens (native 256K) --- ## Surgical Tensor Quantization Map (Audited from GGUF) *The exact tensor breakdown below has been verified directly from the compiled binary weights:* | Layer Group | Sub-Component / Tensor | Qty | Precision | Engineering Rationale | | :--- | :--- | :---: | :---: | :--- | | **Global Output Head** | `output.weight` | 1 | **`Q6_K`** | Preserves near-FP16 token classification; eliminates syntax errors, bracket drops, and hallucinations. | | **Global Embeddings** | `token_embd.weight` | 1 | **`Q4_K`** | High-fidelity vocabulary embedding representation. | | **All Normalizations** | `output_norm`, `attn_*_norm`, `ssm_norm` | 171 | **`F32`** | 100% uncompressed numerical stability across all 40 layers. | | **Expert Routers** | `blk.*.ffn_gate_inp`, `ffn_gate_inp_shexp` | 82 | **`F32`** | 100% uncompressed routing fidelity across 256 micro-experts; zero router drift. | | **Attention Gates** | `blk.*.attn_gate.weight` (Hybrid Layers) | 31 | **`Q8_0`** | High-precision attention gating across hybrid DeltaNet recurrence layers; eliminates crosstalk. | | **Shared Foundation Experts** | `blk.*.ffn_{gate,down,up}_shexp` (All Layers) | 123 | **`Q4_K`** | Foundation knowledge backbone active on 100% of tokens; protected in linear Q4_K for fast streaming. | | **Periodic Full Attention** | `blk.{3,7,11,...}.attn_q/k/v` (10 Anchor Layers) | 30 | **`Q4_K`** | Full quadratic attention anchor checkpoints for deep needle-in-a-haystack retrieval. | | **Periodic Full Attention** | `blk.{3,7,11,...}.attn_output` (10 Anchor Layers) | 10 | **`Q6_K`** | Armored attention output projection over deep context. | | **Recurrent SSM Scales** | `blk.*.ssm_alpha`, `ssm_a`, `ssm_conv1d`, `ssm_dt` | 123 | **`F32`** | Guarded in uncompressed FP32 to prevent DeltaNet recurrent state drift. | | **Linear Attention & SSM** | `blk.*.attn_qkv`, `ssm_beta`, `ssm_out` | 92 | **`Q4_K`** | Linear AVX2 execution; zero SIMD CPU stalls during system RAM streaming. | | **Border MoE Down-Proj** | Layers 0–1 & 38–39 (`ffn_down_exps`) | 4 | **`Q3_K`** | Linear SIMD execution optimized for token entry and exit stability. | | **Border MoE Gate/Up** | Layers 0–1 & 38–39 (`ffn_gate/up_exps`) | 8 | **`IQ3_XXS`** | High-density boundary protection guided by imatrix. | | **Core MoE Down-Proj** | Layers 2–37 (`ffn_down_exps`) | 36 | **`IQ3_XXS`** | Fortified 3.06 bpw residual stream; preserves core mathematical and reasoning capacity. | | **Core MoE Gating** | Layers 2–37 (`ffn_gate_exps`) | 36 | **`IQ2_S`** | High-precision 2.50 bpw SwiGLU gating; eliminates activation noise. | | **Core MoE Up-Proj** | Layers 2–15 (`ffn_up_exps`) | 14 | **`IQ2_S`** | Enhanced 2.50 bpw precision for sensitive early-intermediate feature extraction. | | **Core MoE Up-Proj** | Layers 16–37 (`ffn_up_exps`) | 22 | **`IQ2_XXS`** | Extreme 2.06 bpw compression in deep MoE layers to reach exact 13.0 GB envelope. | | **MTP Companion Head** | `mtp-Qwen3.6-35B-A3B-Q4_0.gguf` | 1 | **`Q4_0`** | Separate low-VRAM speculative draft companion; use with `-md`. | --- ## Hardware Throughput & Offload Benchmarks (RTX 30 / 40 / 50 & RAM Streaming) *Hardware guidance for compatible llama.cpp-style runtimes:* | Hardware Target | Offload Mode | Generation Speed (Est.) | Prompt Prefill Speed (Est.) | Highlights | | :--- | :--- | :---: | :---: | :--- | | **NVIDIA RTX 5080 / 5090 (Blackwell)** | Full GPU (`-ngl 99`) | **approx. 247 – 251 tok/s** | **2,800 – 3,900+ tok/s** | Ultra-high throughput on Blackwell architecture; native 256k window in VRAM | | **NVIDIA RTX 4090 (24GB GDDR6X)** | Full GPU (`-ngl 99`) | **95 – 120+ tok/s** | **2,000 – 2,800+ tok/s** | Linear attention layers slash prefill latency; 256k window in VRAM | | **NVIDIA RTX 4080 (16GB GDDR6X)** | Full GPU (`-ngl 99`) | **75 – 92+ tok/s** | **1,600 – 2,200+ tok/s** | **Entire model fits in VRAM with room for 32k context!** | | **NVIDIA RTX 3090 (24GB GDDR6)** | Full GPU (`-ngl 99`) | **75 – 90+ tok/s** | **1,500 – 2,000+ tok/s** | Full 256k native window in VRAM | | **Workstation / Laptop (DDR4 / DDR5 RAM)** | Hybrid Offload (Few layers in VRAM) | Hardware-dependent | Hardware-dependent | Zero AVX2 CPU stalls; efficient streaming from system RAM | - **Aggressive Hybrid Offload Profile:** Hybrid offload supports reasoning-enabled generation with limited VRAM while the remaining model weights stream from system RAM. > [!NOTE] > ### Empirical Testbed Architecture & Desktop/Server Scaling > - **Empirical Benchmark Hardware:** The hybrid offload and system RAM streaming behavior documented above was measured on a consumer laptop powered by an **Intel 12th Gen Alder Lake architecture featuring a hybrid design of Performance Cores (P-Cores) and Efficient Cores (E-Cores)** paired with dual-channel system RAM and constrained laptop power/thermal envelopes. > - **Thread Scheduling & E-Core Contention:** In hybrid architectures like Alder Lake, OS thread scheduling across background E-Cores and lower single-core mobile power limits introduce memory bandwidth and thread synchronization overhead during CPU dequantization. > - **Dramatic Scaling on Higher-End Processors:** When running on desktop or server processors (such as modern AMD Ryzen 7000 / 9000 Zen 4/5 series or high-TDP Intel desktop platforms with dedicated performance cores, large L3 caches, and high-bandwidth dual- or quad-channel DDR5 running at 6000+ MT/s), **streaming generation speeds and prefill throughput will scale dramatically higher**, substantially exceeding these measured mobile numbers. --- ## The 16GB & 24GB Advantage: Massive Context in VRAM! **Qwen3.6-35B-A3B APEX-I-NanoPlus fits native context on 16GB GPUs and the full 256K window on 24GB GPUs:** | Context Length | Model Weights (Est.) | KV Cache (q8_0, 4 slots) | Compute Buffers | **Total GPU VRAM (Est.)** | Feasibility | | :--- | :--- | :--- | :--- | :--- | :--- | | **32,768 (32k)** | `12.14 GiB` | `0.58 GiB` | `1.80 GiB` | **`14.52 GiB`** | **FULL OFFLOAD ON 16GB GPUs (RTX 4080 / 4070 Ti Super)!** | | **65,536 (64k)** | `12.14 GiB` | `0.92 GiB` | `1.95 GiB` | **`15.01 GiB`** | **FULL OFFLOAD ON 16GB GPUs!** | | **131,072 (128k)**| `12.14 GiB` | `1.58 GiB` | `2.22 GiB` | **`15.94 GiB`** | Full offload on 24GB; tight fit on 16GB | | **262,144 (256k)**| `12.14 GiB` | `2.92 GiB` | `2.80 GiB` | **`17.86 GiB`** | **FULL 256K NATIVE IN 24GB VRAM!** | *Note: Leaves comfortable headroom for display drivers and compute buffers on standard 16GB and 24GB GPUs.* > [!TIP] > ### πŸ’‘ 16GB GPU Verification (Single Stream / Desktop) > Single-stream desktop inference with FlashAttention consumes **approx. 14.5 GB at 32,768 ctx** and **approx. 15.0 GB at 65,536 ctx** (`-ctk q8_0 -ctv q8_0 -fa on`). This confirms that **32K to 64K context runs completely in VRAM on 16GB cards (RTX 4080 / RTX 4070 Ti Super / RTX 5080)** without system RAM offload! --- ## Recommended Configuration & Setup ### 1. `llama-cli` with MTP Speculative Decoding & Vision Projector ```bash llama-cli \ -m Qwen3.6-35B-A3B.APEX-I-NanoPlus.gguf \ -md mtp-Qwen3.6-35B-A3B-Q4_0.gguf \ --mmproj mmproj-Q8_0.gguf \ -ngl 99 \ -c 32768 \ --temp 0.6 \ -p "Explain the implementation of a thread-safe LRU cache in Rust." ``` ### 2. High-Performance OpenAI-Compatible Server ```bash llama-server.exe \ -m Qwen3.6-35B-A3B.APEX-I-NanoPlus.gguf \ -md mtp-Qwen3.6-35B-A3B-Q4_0.gguf \ --mmproj mmproj-Q8_0.gguf \ -ngl 99 \ --port 8080 \ --parallel 4 \ --flash-attn on \ --fit on \ -c 65536 \ --cache-type-k q8_0 \ --cache-type-v q8_0 ``` --- ### βš™οΈ Recommended Generation Parameters (Qwen Official) Sampling metadata recorded from [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) in the completed build: | Hyperparameter | Value | Description / Creator Notice | | :--- | :---: | :--- | | **Temperature** | `0.60` | Recommended for reasoning and coding stability. | | **Top-P** | `0.95` | Source GGUF sampling metadata. | | **Top-K** | `20` | Source GGUF sampling metadata. | | **Max New Tokens** | Runtime-dependent | Select for the target workload. | --- > [!IMPORTANT] > > ### πŸ” Model Inherent Behavior vs. Quantization Fidelity Notice > Any behavioral nuances, stylistic tendencies, domain-specific habits, or zero-shot edge-case oversights **stem entirely from the original unquantized checkpoint weights and fine-tuning distribution, NOT from the APEX-I quantization process.** > Handcrafted APEX-I-NanoPlus strictly preserves mathematical tensor fidelityβ€”keeping 100% of expert routing matrices (`gate_inp`) in uncompressed `F32` (zero router drift), armoring the token output head in `Q6_K`, safeguarding attention gates in `Q8_0`, and protecting the down-projection residual stream in `IQ3_XXS`. Empirical verification records the final GGUF perplexity at **5.5244 Β± 0.12916**. > [!IMPORTANT] > ### CRITICAL ADVISORY FOR CODING WORKFLOWS: PREVENTING SYNTAX & TOKEN SWAPPING > In programming code, brackets (`{`, `}`), assignment operators (`=`), and indentation whitespace repeat constantly across multi-line structures. > > **Common Issue:** Many local frontends (such as LM Studio defaults, Ollama, or web interfaces) ship with `repeat_penalty` set to `1.1` or `1.15`. While this prevents loops in creative prose, applying repeat penalties to code artificially penalizes necessary syntax tokens. When the logit of `{` drops, the model is forced to emit the next closest mathematical token (`=` or `[`), resulting in character swapping or dropped/doubled whitespace. > > **Verified Upstream Behavior:** This self-correcting behavior (where the model notices the mistake in its thinking loop but repeats the substitution) is documented on official upstream base checkpoints and Q8 builds ([Ornith Discussion #34](https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B/discussions/34) and [Discussion #22](https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF/discussions/22)). It is completely eliminated by proper sampling configuration: > > 1. **Disable Repeat Penalties (Required for Code):** > - `repeat_penalty: 1.0` (strictly disabled) > - `presence_penalty: 0.0` > - `frequency_penalty: 0.0` > 2. **Calibrate Samplers:** > - `temperature: 0.60` (or `0.20` - `0.30` for strict, deterministic code syntax) > - `min_p: 0.05` (prunes low-probability noise tokens effectively) > - `top_p: 0.95` > - `top_k: 20` > 3. **Native Jinja Formatting:** Always pass the `--jinja` flag so the tokenizer handles leading-space BPE tokens (` {` vs `{`, ` =` vs `=`) cleanly. > [!TIP] > ### HARDENED AGENTIC CHAT TEMPLATE (JINJA) > An optimized `chat_template.jinja` is included at the root of this repository. It hardens agent workflows and multi-turn stability: > > 1. **Native `reasoning_effort` Multi-Level Control:** > - `low` / `minimal`: Keeps internal thinking concise and focused strictly on immediate execution steps to minimize latency in automated loops. > - `medium` (default): Balanced, structured reasoning process with standard analytical depth. > - `high` / `xhigh`: Guides the model to formulate a clear implementation plan upfront before generating code, avoiding circular self-doubt loops. > - `none` / `off`: Closes the thinking block immediately (`\n\n`) when reasoning is disabled. > 2. **Tool-Calling Safeguard (Anti-Premature Stop):** Prevents the model from terminating a turn (`<|im_end|>`) at a colon or action declaration prior to outputting ``. > 3. **Multi-Turn Thinking Memory:** Preserves historical `` blocks across turns by default, preventing context distribution drift in 78K+ token runs. > > **Usage with llama-server:** > ```bash > llama-server -m Model.gguf --chat-template-file chat_template.jinja --reasoning-effort medium > ``` ## Optional Support Gold Ship dancing If these MiniPlus or NanoPlus releases have been useful to you and you would like to support the work, you can do so voluntarily through https://ko-fi.com/isvalorum. Your contribution helps with evaluation, hosting, and future handcrafted quantizations. Every release will always remain free to download and use; there are no paywalled files, updates, or features.