Upload README.md with huggingface_hub

README.md

@@ -16,6 +16,8 @@ tags:
   - ollama
   - batiai
   - on-device
+  - agentic
+  - coding
 base_model: Qwen/Qwen3.6-35B-A3B
 pipeline_tag: text-generation
 library_name: llama.cpp
@@ -26,11 +28,14 @@ library_name: llama.cpp
 <p align="center">
   <a href="https://flow.bati.ai"><img src="https://img.shields.io/badge/BatiFlow-macOS%20AI%20Automation-blue?style=for-the-badge&logo=apple" alt="BatiFlow"></a>
   <a href="https://ollama.com/batiai/qwen3.6-35b"><img src="https://img.shields.io/badge/Ollama-batiai%2Fqwen3.6--35b-green?style=for-the-badge" alt="Ollama"></a>
+  <a href="https://huggingface.co/Qwen/Qwen3.6-35B-A3B"><img src="https://img.shields.io/badge/Upstream-Qwen3.6--35B--A3B-orange?style=for-the-badge" alt="Upstream"></a>
 </p>
 
-> imatrix-calibrated GGUF quantizations of **Qwen/Qwen3.6-35B-A3B** (text-only) for on-device AI on Mac.
+> **"Agentic Coding Power, Now Open to All"** — imatrix-calibrated GGUF quantizations of **Qwen/Qwen3.6-35B-A3B** (text-only) for on-device AI on Mac.
 > Built and verified by [BatiAI](https://bati.ai) for [BatiFlow](https://flow.bati.ai) — free, unlimited, on-device AI automation.
 
+Released by Alibaba on **April 15, 2026** as the successor to Qwen 3.5 35B-A3B, with substantial upgrades in **agentic coding**, **frontend workflows**, **repository-level reasoning**, and **thinking preservation** for iterative development.
+
 ## Quick Start
 
 ```bash
@@ -43,7 +48,7 @@ ollama pull batiai/qwen3.6-35b:iq4
 ollama run batiai/qwen3.6-35b:iq4
 ```
 
-Aliases `:q3` and `:q4` point to the same blobs.
+Aliases `:q3` / `:q4` point to the same blobs as `:iq3` / `:iq4`.
 
 ## Available Quantizations
 
@@ -52,17 +57,61 @@ Aliases `:q3` and `:q4` point to the same blobs.
 | `:iq3` / `:q3` | **IQ3_XXS** (imatrix) | **13 GB** | 16 GB | Mac mini / MacBook Air 16GB |
 | `:iq4` / `:q4` | **IQ4_XS** (imatrix) | **18 GB** | 24 GB | MacBook Pro / Mac Studio 24GB+ |
 
-Both quants use an **importance matrix** computed from wikitext-2-raw to preserve quality at low bit-widths. IQ quants generally outperform K-quants of similar size.
+Both quants use an **importance matrix** computed from wikitext-2-raw to preserve quality at low bit-widths — IQ quants generally outperform K-quants of similar size.
 
 ## Why Qwen 3.6 35B-A3B?
 
-- **MoE — 35B total, only ~3B active per token** → 9× less compute than a 27B Dense model
-- **262K native context** (1M with YaRN scaling)
-- **Gated DeltaNet hybrid attention** — long-context efficiency
-- **Function calling** — `qwen3_coder` parser (vLLM/SGLang)
-- **Agentic workflows** — tuned for repository-level coding, multi-tool reasoning
-- **Multilingual** — Korean / Japanese / Chinese / English
-- **Apache 2.0** — commercial-friendly
+Upstream headline: _"Agentic Coding Power, Now Open to All"_ — the model is tuned for multi-step coding agents, long-horizon repo reasoning, and tool use.
+
+### Benchmarks (official Qwen BF16 figures)
+
+**Coding & Agentic**
+
+| Benchmark | Qwen 3.6-35B-A3B | Qwen 3.5-35B-A3B | Gemma 4-31B |
+|---|---:|---:|---:|
+| SWE-bench Verified | **73.4** | 70.0 | 52.0 |
+| SWE-bench Multilingual | **67.2** | — | 51.7 |
+| SWE-bench Pro | **49.5** | — | 35.7 |
+| Terminal-Bench 2.0 | **51.5** | 40.5 | 42.9 |
+| QwenWebBench | **1397** | 978 | — |
+
+**Math & Reasoning**
+
+| Benchmark | Qwen 3.6-35B-A3B | Gemma 4-31B |
+|---|---:|---:|
+| AIME26 | **92.7** | 89.2 |
+| GPQA | **86.0** | — |
+| HMMT Feb 26 | **83.6** | — |
+| HLE | **21.4** | — |
+| LiveCodeBench v6 | **80.4** | — |
+
+**General Knowledge**
+
+| Benchmark | Qwen 3.6-35B-A3B |
+|---|---:|
+| MMLU-Pro | 85.2 |
+| MMLU-Redux | 93.3 |
+| SuperGPQA | 64.7 |
+| C-Eval | 90.0 |
+
+**Agent / Tool Use**
+
+| Benchmark | Qwen 3.6-35B-A3B |
+|---|---:|
+| TAU3-Bench | 67.2 |
+| MCP-Atlas | 62.8 |
+| WideSearch | 60.1 |
+| MCPMark | 37.0 |
+| Tool Decathlon | 26.9 |
+
+### Key takeaways
+
+- **SWE-bench Verified jumps +3.4** over Qwen 3.5 to **73.4** — top-tier agentic-coding among open models
+- **Terminal-Bench 2.0 +11.0** over 3.5 → genuine real-world command-line competence
+- **QwenWebBench 1397** vs 978 for 3.5 — a **43% jump** in agentic web tasks
+- **Beats Gemma 4-31B on every published coding & reasoning benchmark** despite Gemma being a similar-sized dense model (A3B only activates 3B params per token)
+
+_Note: these are upstream BF16 figures. IQ3_XXS / IQ4_XS quantization may cost a few points on the hardest benchmarks — post your own bench results and we'll update this card._
 
 ## MoE Advantage
 
@@ -71,10 +120,10 @@ Both quants use an **importance matrix** computed from wikitext-2-raw to preserv
 | Total params | 35B | 27B |
 | **Active params / token** | **3B** | **27B** |
 | Experts | 256 (8 routed + 1 shared) | — |
-| Typical VRAM | ~23 GB (IQ4) | ~28 GB |
+| Typical VRAM (IQ4) | ~23 GB | ~28 GB |
 | Relative speed | **Faster** | Baseline |
 
-Only 9 of 256 experts fire per token — same reasoning, far less compute.
+Only 9 of 256 experts fire per token — same reasoning capacity, far less compute.
 
 ## RAM Requirements (on-device)
 
@@ -85,26 +134,37 @@ Only 9 of 256 experts fire per token — same reasoning, far less compute.
 | 32 GB        | ✅              | ✅ comfortable  |
 | 48 GB+       | ✅              | ✅ ideal        |
 
-## Benchmarks
+## On-device Benchmarks (measured)
 
-Benchmarks on Apple Silicon will be updated as measurements come in. To reproduce on your Mac:
+Measured tokens/s on your Mac — update this card with your own numbers:
 
 ```bash
 ollama run batiai/qwen3.6-35b:iq4 --verbose "Write a haiku about Seoul in autumn."
 ```
 
-The `--verbose` flag prints prompt-eval rate, token-generation rate, and memory usage.
+| Mac | IQ3_XXS | IQ4_XS |
+|-----|--------:|-------:|
+| _pending community measurements_ | — | — |
 
 ## Note on the "3.6" Naming
 
-Upstream Qwen released this model as **Qwen 3.6** publicly. Internally the Hugging Face config still registers the architecture as `Qwen3_5MoeForConditionalGeneration` (a transitional class name carried over from the 3.5 line). llama.cpp handles this class via its `Qwen3_5MoeTextModel` converter, which is what these GGUFs were built from — text-only, vision tower excluded.
+Upstream Qwen released this model as **Qwen 3.6** publicly. Internally the Hugging Face config still registers the architecture as `Qwen3_5MoeForConditionalGeneration` (a transitional class name carried over from the 3.5 line). llama.cpp handles this via its `Qwen3_5MoeTextModel` converter, which is what these GGUFs were built from — **text-only, vision tower excluded**. For the vision-language capabilities (MMMU 81.7, MathVista 86.4, etc.), use the upstream multimodal weights directly.
 
 ## Technical Details
 
 - **Original Model**: [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B)
-- **Architecture**: MoE + Gated DeltaNet hybrid (35B total, ~3B active, 40 layers, hidden 2048)
+- **Released**: 2026-04-15
+- **Architecture**: MoE + **Gated DeltaNet hybrid attention**
+  - 40 layers, hidden 2048, expert-intermediate 512
+  - Layout: 10× (3× Gated DeltaNet → MoE + 1× Gated Attention → MoE)
+  - Linear-attention heads: 32 V / 16 QK (head dim 128)
+  - Softmax-attention heads: 16 Q / 2 KV (head dim 256, RoPE dim 64)
+- **Parameters**: 35 B total, ~3 B active per forward pass
 - **Experts**: 256 total, 8 routed + 1 shared per token
-- **Context Window**: 262,144 tokens native (extensible to ~1M via YaRN)
+- **Context Window**: 262,144 tokens native (extensible to ~1,010,000 via YaRN)
+- **Vocabulary**: 248,320 tokens (padded)
+- **Training**: Multi-token Prediction (MTP) applied for speculative decoding
+- **Modes**: thinking / non-thinking switchable
 - **License**: Apache 2.0
 - **Quantized with**: [llama.cpp](https://github.com/ggml-org/llama.cpp) build `bafae2765`
 - **Quantized by**: [BatiAI](https://bati.ai)
@@ -144,3 +204,7 @@ No third-party intermediaries. Direct from official Qwen weights.
 This repo mirrors the upstream license. `Qwen/Qwen3.6-35B-A3B` is released under **Apache 2.0** — commercial use permitted.
 
 BatiAI's quantization pipeline is MIT.
+
+## Sources
+
+Benchmark numbers in this card come from the official upstream [Qwen/Qwen3.6-35B-A3B](https://huggingface.co/Qwen/Qwen3.6-35B-A3B) model card and Qwen's [research blog](https://qwen.ai/blog?id=qwen3.6-35b-a3b). Quantization and on-device numbers are measured by BatiAI.