Update README.md

README.md

@@ -3,19 +3,22 @@ license: apache-2.0
 base_model:
 - Qwen/Qwen3.8-27B
 tags:
-  - gguf
-  - llama.cpp
-  - quantization
-  - mixed-precision
-  - iq2
-  - iq3
-  - qwen
+- gguf
+- llama.cpp
+- quantization
+- mixed-precision
+- iq2
+- iq3
+- qwen
+pipeline_tag: image-text-to-text
 ---
 
-# TL; DR
-
-Right now the best model that I have tested is the **IQ2_E_S** in overall **quality** and **quality/size ratio** (yes, better performance in general than the Q3). I'll keep testing the performance on the **IQ2_E_S and IQ3_S_S** that are the Two models that works.
+# TL; DR Updated (14/09/2026)
+Currently, the standout quant is **IQ3_E_S_V2**, offering the best balance of **overall quality** and **quality-to-size ratio**. In the ultra-low bitrate tier (~7.5 GB / sub-8 GB), I am actively benchmarking **IQ2_E_S_V2** and **IQ2_E_S**, which proved to be the most viable variants:
 
+* **IQ3_E_S_V2** — **Best Overall**: Highly capable in long-context coding and complex multi-step reasoning. While it experiences occasional minor attention slips or hallucinations, it remains significantly more reliable and consistent than the lower-bit variants.
+* **IQ2_E_S_V2** — **Best for Code & Agentic Workflows (~2.2 BPW)**: Remarkably capable in long-context tasks, code synthesis, and architectural understanding (e.g., capable of generating fully playable interactive scripts in a single prompt). However, it is prone to repetition loops in standard conversational text, though it typically self-terminates without freezing. *(Tip: Setting `repeat_penalty: 1.15`–`1.20` and `presence_penalty: 0.2` effectively eliminates these loops).*
+* **IQ2_E_S** — **Best Stability for Direct Q&A (~2.2 BPW)**: Shows strong conversational stability, strict output formatting, and high resistance to counterfactual traps on isolated prompts. However, it suffers from severe attention degradation during complex logic execution, making it unsuitable for long-context generation or functional coding tasks.
 
 # Qwen3.8-27B GGUF Quantizations (IQ2_E_S & IQ3_E_S)
 
@@ -29,80 +32,6 @@ Quantizing a 27B model down to this range yields two distinct results:
 1. **IQ2_E_S (2.40 BPW, ~7.51 GiB):** Fits tightly into an 8 GB VRAM budget. While low-bit quantization introduces inevitable degradation, perplexity remains bounded, making the model practically usable for lighter or structured generation tasks.
 2. **IQ3_E_S (3.00 BPW, ~9.41 GiB):** A balanced quantization targeting 10–12 GB GPUs, showing theoretical perplexity retention close to the unquantized baseline.
 
----
-
-## Perplexity Evaluation
-
-Perplexity was evaluated using standard Wiki-style test sets and compared against the FP16/BF16 reference as well as external community baselines.
-
-| Candidate | Size | BPW | Wiki-style PPL | Notes |
-| :--- | :---: | :---: | :---: | :--- |
-| **BF16 GGUF** | 50.89 GiB | 16.00 | 7.15 ± 0.12 | Unquantized Baseline |
-| **IQ3_E_S** *(Ours)* | 9.41 GiB | 3.00 | **7.4453 ± 0.1287** | Update: PPL score doesn't show real performance. See real test usage bellow |
-| `empero-ai/Qwen3.8-27B-Ridge-GGUF` | 11.73 GiB | 3.69 | 7.82 ± 0.14 | External Reference |
-| **IQ3_S_S** *(Ours)* | 9.41 GiB | 3.00 | **7.9560 +/- 0.1382** | Best balance (10–12 GB VRAM) |
-| **IQ2_E_S** *(Ours)* | 7.51 GiB | 2.40 | **8.8587 ± 0.1576** | Sub-8GB Footprint |
-
-*Update: The IQ3_S_S quantization shows higher performance level based on a Gemini assisted benchmark.*
-
-
-
-# **Benchmark Report: Low-Bitrate Evaluation of Qwen-27B (~9 GB / ~3 BPW)**  
-This evaluation compares four aggressive quantization strategies applied to the 27B parameter base model, all targeting a footprint of ~9.4 GB (~2.5–3.0 BPW). Testing was executed locally using llama-server without reasoning/chain-of-thought active, forcing single-pass forward inference to reveal structural tensor damage.  
-## **1. Evaluated Models & Configurations**  
-| | | | |  
-|-|-|-|-|  
-| **Model Identifier** | **Source / Quantization Strategy** | **Target Size** | **Score (0–10)** |   
-| **Quantization 1** | Custom Hybrid 3.0 BPW (IQ3_S_S) | ~9.41 GB | **7.5 / 10** |   
-| **Quantization 2** | Custom Hybrid 3.0 BPW (IQ3_E_S) | ~9.41 GB | **3.0 / 10** |   
-| **Quantization 3** | Unsloth UD-Q2_K_XL | ~9.1 GB | **1.0 / 10** |   
-| **Quantization 4** | ISTA-DASLab GSQ-RCO-GGUF (IQ2_S) | ~8.62 GB | **8.5 / 10** |   
-## **2. Test Battery & Evaluation**  
-### **Test 1: Conceptual Premise & Negation Handling**  
-- **Prompt:** *"Who discovered Christopher Columbus?"*  
-- **Objective:** Detect the false premise (Columbus discovered America, he was not discovered).  
-- **Results:**  
-  - **Quantization 1:** Clean structural breakdown; identified the premise flaw immediately and offered historically sound alternative interpretations.  
-  - **Quantization 2:** Tautological degradation (*"Christopher Columbus, also known as Christopher Columbus"*) and historical hallucinations ( *"appointed king by Charles V"*).  
-  - **Quantization 3:** Complete control-token breakdown. Triggered an unsolicited internal tool call and defaulted to a generic greeting loop in English.  
-  - **Quantization 4:** Flawless detection. Accurately cited period historians (*Bartolomé de las Casas*,  *Gonzalo Fernández de Oviedo*).  
-### **Test 2: Sequential State Tracking & Working Memory**  
-- **Task:** Track discrete movements of three colored tokens across three containers over four sequential mutation steps.  
-- **Ground Truth:** Box A: [blue], Box B: [empty], Box C: [red].  
-- **Results:**  
-  - **Quantization 1:** Froze at Step 1 (A: red, B: green, C: blue). Failed subsequent temporal transitions.  
-  - **Quantization 2:** Cloned state tokens (B: red, C: red), violating physical conservation constraints.  
-  - **Quantization 3:** Excluded due to system-level instability.  
-  - **Quantization 4:** Successfully tracked states through Step 3 (A: blue, B: green, C: red), failing only the final step (discarding green).  
-### **Test 3: Static Code Debugging (Binary Search Off-By-One)**  
-- **Task:** Identify the line and failure mode of izq = medio in a Python binary search implementation.  
-- **Results:**  
-  - **Quantization 1:** Correctly flagged the infinite loop condition and provided the exact standard patch (izq = medio + 1) with clean code.  
-  - **Quantization 2:** Identified the infinite loop, but provided minimal context.  
-  - **Quantization 4:** Top-tier output. Pinpointed the exact line number, explained the edge case, and generated a reproducible counterexample (arr = [1, 2], target = 3).  
-### **Test 4: Counterfactual Resistance**  
-- **Task:** Name the first human on the Moon under the counterfactual premise that Apollo 11 failed and no human landed in the 20th century.  
-- **Results:**  
-  - **Quantization 1 & 4:** Both fell into secondary token attractor wells, attributing the landing to *Buzz Aldrin*, though Q4 explicitly broke the prompt constraint by citing Apollo 11.  
-  - **Quantization 2:** Severe hallucination (*"Buzz Aldrin in 1972 on Apollo 12"*).  
-### **Test 5: Language Consistency & Python Reference Semantics**  
-- **Task:** Identify the bug in def registrar_evento(evento, metadata={}) and supply a one-line idiom fix.  
-- **Results:**  
-  - **Quantization 1:** Minor semantic slip (called the mutable argument a "list" initially), but provided the correct dictionary fix and inline patch.  
-  - **Quantization 4:** Impeccable technical accuracy. Precisely identified mutable default state sharing across calls and provided the idiomatic metadata=None pattern.  
-## **3. Tool-Use Artifacts in llama-server**  
-At low bitrates (< 3.0 BPW), logits associated with special function-calling tokens (<tool_call>) experience variance shifts. When faced with high internal uncertainty (e.g., token-counting constraints or strict negation), **Quantization 1** and  **Quantization 4** occasionally emitted false-positive tool calls to default server utilities (system inspection / system clock).  
-This is a known calibration artifact of low-bit quantization on models fine-tuned with function calling. It can be mitigated at runtime by disabling function-calling schemas in inference endpoints where external tools are not required.  
-## **4. Final Verdict & Key Takeaways**  
-- **ISTA-DASLab (Quantization 4):** The overall benchmark winner. Second-order optimization and coordinate rotations (GSQ/RCO) prevent weight outliers from crushing key attention blocks at 2-bit scales, preserving fine-grained algorithmic reasoning and line-accurate code debugging.  
-- **Custom Hybrid 3.0 BPW - ** **IQ3_S_S** ** (Quantization 1):** Strong performer. Retains syntax integrity, structural formatting, and standard Python conventions far better than standard community low-bit builds. It represents a viable sub-10 GB daily driver for coding and conversational flows.  
-- **Custom Hybrid 3.0 BPW - ** **IQ3_** **E** **_S** ** (Quantization 2):** Unviable. Shows severe activation collapse in semantic layers, leading to loops, factual drift, and entity hallucinations.  
-- **Unsloth UD-Q2_K_XL (Quantization 3):** Discarded. Suffers critical corruption on chat/system control tokens, resulting in unsolicited function calls and output generation aborts.
-
-
-
-
----
   
   ## Backend & Hardware Compatibility Notes (Intel Arc / oneAPI)