Update private release-card draft v2

README.md

@@ -14,59 +14,87 @@ tags:
 
 # Qwen3.5 4B — TheStageAI GGUF
 
-Four text-only GGUF checkpoints from **1.519 GB** to **4.486 GB**, evaluated on IFEval and MMLU-Pro where a complete release score is available.
+Four optimized language-model GGUF files for text-only deployment with llama.cpp.
 
-[Qwen3.5 0.8B](https://huggingface.co/TheStageAI/Qwen3.5-0.8B-GGUF) · [Qwen3.5 2B](https://huggingface.co/TheStageAI/Qwen3.5-2B-GGUF) · **[Qwen3.5 4B](https://huggingface.co/TheStageAI/Qwen3.5-4B-GGUF)** · [Qwen3.5 9B](https://huggingface.co/TheStageAI/Qwen3.5-9B-GGUF) · [Gemma 4 E2B IT](https://huggingface.co/TheStageAI/gemma-4-E2B-it-GGUF) · [Gemma 4 E4B IT](https://huggingface.co/TheStageAI/gemma-4-E4B-it-GGUF) · [Gemma 4 12B IT](https://huggingface.co/TheStageAI/gemma-4-12B-it-GGUF)
+> **Recommended: M · 2.39 GB** — **78.86 MMLU-Pro**, 99.1% of the BF16 reference.
 
-## Choose a file
+**Qwen 3.5 family:** [0.8B](https://huggingface.co/TheStageAI/Qwen3.5-0.8B-GGUF) · [2B](https://huggingface.co/TheStageAI/Qwen3.5-2B-GGUF) · **[4B](https://huggingface.co/TheStageAI/Qwen3.5-4B-GGUF)** · [9B](https://huggingface.co/TheStageAI/Qwen3.5-9B-GGUF)
 
-| Variant | Hub class | File | Size | Whole-file BPW | IFEval P / I (%) | MMLU-Pro (%) | Use |
-| --- | --- | --- | ---: | ---: | ---: | ---: | --- |
-| XS | `Q3_K_S` | `Qwen3.5-4B-XS-TS-Q3_K_S.gguf` | 1.519 GB | 2.889 | 70.43 / 78.30 | — | Minimum size |
-| S | `Q4_K_S` | `Qwen3.5-4B-S-TS-Q4_K_S.gguf` | 1.904 GB | 3.622 | 77.82 / 83.93 | 74.39 | Compact |
-| **M** | `Q4_K_M` | `Qwen3.5-4B-M-TS-Q4_K_M.gguf` | 2.386 GB | 4.538 | 80.22 / 86.09 | 78.86 | **Recommended · Uniform Q4** |
-| L | `Q8_0` | `Qwen3.5-4B-L-TS-Q8_0.gguf` | 4.486 GB | 8.533 | 81.70 / 87.05 | 79.59 | Uniform Q8 |
+## Choose a checkpoint
 
-> **Hub class versus file type:** XS and S are TheStage mixed-precision schedules. Their Hub labels describe the whole-file size class for discoverability; they are not stock `Q3_K_S`, `Q3_K_L`, or `Q4_K_S` conversions. M and L use the actual `MOSTLY_Q4_K_M` and `MOSTLY_Q8_0` GGUF file types. Exact tensor-type counts and SHA-256 digests are in [`release-manifest.json`](./release-manifest.json).
+| Variant | Size | Best for | File |
+| --- | ---: | --- | --- |
+| XS | 1.52 GB | Minimum footprint · non-thinking only | [GGUF](./Qwen3.5-4B-XS-TS-Q3_K_S.gguf) |
+| S | 1.90 GB | Compact | [GGUF](./Qwen3.5-4B-S-TS-Q4_K_S.gguf) |
+| **M** | **2.39 GB** | **Recommended · Q4 reference** | [GGUF](./Qwen3.5-4B-M-TS-Q4_K_M.gguf) |
+| L | 4.49 GB | Maximum fidelity | [GGUF](./Qwen3.5-4B-L-TS-Q8_0.gguf) |
 
-IFEval is shown as prompt-strict / instruction-strict. The BF16 reference scored **82.44 / 87.53** on IFEval and **79.55** on MMLU-Pro.
+Sizes use decimal MB/GB. Exact byte counts are recorded in [`release-manifest.json`](./release-manifest.json).
 
 ## Quickstart
 
-Use a current [llama.cpp](https://github.com/ggml-org/llama.cpp) build. The command below selects the recommended **M** file explicitly through its Hub quant label:
+The command below selects the recommended **M** checkpoint by exact filename:
 
 ```bash
-llama-cli -hf TheStageAI/Qwen3.5-4B-GGUF:Q4_K_M
+llama-cli \
+  --hf-repo TheStageAI/Qwen3.5-4B-GGUF \
+  --hf-file Qwen3.5-4B-M-TS-Q4_K_M.gguf
 ```
 
-## Evaluation
+## Quality
+
+| Variant | IFEval prompt / instruction strict (%) | MMLU-Pro (%) |
+| --- | ---: | ---: |
+| BF16 reference | 82.44 / 87.53 | 79.55 |
+| XS | 70.43 / 78.30 | — |
+| S | 77.82 / 83.93 | 74.39 |
+| **M** | 80.22 / 86.09 | 78.86 |
+| L | 81.70 / 87.05 | 79.59 |
 
 - **IFEval:** 541 prompts, native chat template, `enable_thinking=false`, temperature 0.
-- **MMLU-Pro:** 12,032 questions for complete rows, native chat template, `enable_thinking=true`, temperature 1, top-p 0.95.
-- A dash means **not reported**, not zero and not a score reconstructed from partial subjects.
+- **MMLU-Pro:** complete 12,032-question runs only, native chat template, `enable_thinking=true`, temperature 1, top-p 0.95, 32,768-token output limit.
+- `—` means that no complete headline MMLU-Pro run is reported. Partial-subject accuracies are not used as release scores.
+
+The benchmarks intentionally exercise different modes: IFEval measures non-thinking instruction and format adherence, while MMLU-Pro includes long sampled reasoning.
 
-These protocols intentionally exercise different operating modes. IFEval measures non-thinking instruction and format adherence; MMLU-Pro includes long sampled reasoning.
+## Reasoning mode
 
-## What this release is
+**XS is not a supported reasoning operating point.** It is released for minimum-footprint, non-thinking chat and instruction following. Use S, M, or L when reasoning quality matters. In Qwen XS, long-thinking evaluation produced unstable trajectories, substantially longer generations, and frequent 32,768-token output-limit hits. Run XS with `--reasoning off`. The XS MMLU-Pro cell is therefore marked `—` (N/A).
 
-This release starts from [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B) at revision [`851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a`](https://huggingface.co/Qwen/Qwen3.5-4B/tree/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a) and applies TheStageAI's scale-tuned deployment compression pipeline.
+## About this release
 
-- **XS:** minimum-size adaptive mixed precision.
+This release starts from [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B) at revision [`851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a`](https://huggingface.co/Qwen/Qwen3.5-4B/tree/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a) and applies TheStageAI's adaptive per-tensor precision assignment with tuned quantization scales.
+
+- **XS:** minimum-footprint adaptive mixed precision for non-thinking use.
 - **S:** compact adaptive mixed precision.
-- **M:** uniform Q4 operating point.
-- **L:** uniform Q8 operating point.
+- **M:** uniform Q4 tensor assignment and reference operating point.
+- **L:** uniform Q8_0 tensor assignment for maximum fidelity.
+
+The recommendation is selected separately for each base model from the release evaluations.
+
+<details>
+<summary><b>Quantization labels and technical details</b></summary>
+
+| Variant | Hub selector | GGUF file type | Whole-file BPW |
+| --- | --- | --- | ---: |
+| XS | `Q3_K_S` | `MOSTLY_Q2_K` | 2.889 |
+| S | `Q4_K_S` | `MOSTLY_Q2_K` | 3.622 |
+| M | `Q4_K_M` | `MOSTLY_Q4_K_M` | 4.538 |
+| L | `Q8_0` | `MOSTLY_Q8_0` | 8.533 |
+
+XS and S are TheStage mixed-precision schedules. Their public Q-type suffixes provide an approximate size class for Hub discoverability; exact tensor-type inventories are recorded in [`release-manifest.json`](./release-manifest.json). M and L use uniform Q4_K and Q8_0 assignments for quantized decoder tensors.
 
-The four products are operating points, not a promise that benchmark scores increase monotonically with file size. The recommended row is selected separately for each base model from the release evaluations.
+</details>
 
 ## Limitations
 
-- The files in this repository contain the language-model GGUF. Multimodal projector files are not included.
+- These are language-model GGUF files for text-only deployment. Multimodal projector files are not included.
+- File size does not include KV cache or runtime buffers; total memory use grows with context length.
 - Small score reversals between BF16 and quantized rows should be read as evaluation variation, not as a claim that quantization improves the base model.
-- For XS, use `enable_thinking=false`. A headline thinking-mode MMLU-Pro score is not reported when long generations do not produce a stable product metric.
 
 ## Provenance
 
-The exact base revision, file sizes, GGUF file types, whole-file BPW, SHA-256 digests, held-out KL values, and evaluation IDs are recorded in [`release-manifest.json`](./release-manifest.json). Release filenames differ from the internal artifact paths; the payload bytes must match those digests exactly.
+The exact base revision, byte sizes, GGUF file types, whole-file BPW, tensor inventories, SHA-256 digests, held-out KL values, and evaluation IDs are recorded in [`release-manifest.json`](./release-manifest.json). Use its SHA-256 digests to verify downloaded files.
 
 ## License
 

release-manifest.json

@@ -28,6 +28,7 @@
     "mmlu_pro": {
       "accuracy": 0.7954621010638298,
       "correct": 9571,
+      "not_reported_reason": null,
       "run_id": "release_20260715_bf16_qwen35_mmlu_pro_cardlike_v2_qwen3p5_4b_bf16_mmlu_pro_official",
       "sample_count": 12032,
       "status": "complete"
@@ -35,12 +36,20 @@
   },
   "display_name": "Qwen3.5 4B",
   "family": "Qwen 3.5",
-  "generated_at": "2026-07-20T21:32:16.184050+00:00",
+  "generated_at": "2026-07-20T22:39:04.914041+00:00",
   "license": "apache-2.0",
   "model_key": "qwen3p5_4b",
+  "reasoning_policy": {
+    "reasoning_products": [
+      "S",
+      "M",
+      "L"
+    ],
+    "xs": "non_thinking_only"
+  },
   "recommended_product": "M",
   "repo_id": "TheStageAI/Qwen3.5-4B-GGUF",
-  "schema": "thestageai.gguf_release_manifest_v1",
+  "schema": "thestageai.gguf_release_manifest_v2",
   "variants": [
     {
       "benchmarks": {
@@ -56,6 +65,7 @@
         "mmlu_pro": {
           "accuracy": null,
           "correct": null,
+          "not_reported_reason": "unstable_long_thinking_behavior",
           "run_id": null,
           "sample_count": null,
           "status": "not_reported"
@@ -71,8 +81,9 @@
       "lm_head_policy": "skip_tied",
       "output_embedding_mode": "tied_alias",
       "parameter_count": 4205751296,
-      "positioning": "Minimum size",
+      "positioning": "Minimum footprint \u00b7 non-thinking only",
       "product": "XS",
+      "reasoning_support": "non_thinking_only",
       "recommended": false,
       "sha256": "a1d49f1d782073748589d0df0893658040231866ae6d69667994bef5cc6ce61f",
       "tensor_count": 426,
@@ -99,6 +110,7 @@
         "mmlu_pro": {
           "accuracy": 0.7439328457446809,
           "correct": 8951,
+          "not_reported_reason": null,
           "run_id": "release_20260715_quant_qwen35_mmlu_pro_cardlike_v2_qwen3p5_4b_s_mmlu_pro_official",
           "sample_count": 12032,
           "status": "complete"
@@ -116,6 +128,7 @@
       "parameter_count": 4205751296,
       "positioning": "Compact",
       "product": "S",
+      "reasoning_support": "supported",
       "recommended": false,
       "sha256": "9f020bfc77a9503a641e3e0511de04b4bfba16b9facfa62c3772b4f8e318ba55",
       "tensor_count": 426,
@@ -144,6 +157,7 @@
         "mmlu_pro": {
           "accuracy": 0.7886469414893617,
           "correct": 9489,
+          "not_reported_reason": null,
           "run_id": "release_20260715_quant_qwen35_mmlu_pro_cardlike_v2_qwen3p5_4b_m_mmlu_pro_official",
           "sample_count": 12032,
           "status": "complete"
@@ -159,8 +173,9 @@
       "lm_head_policy": "skip_tied",
       "output_embedding_mode": "tied_alias",
       "parameter_count": 4205751296,
-      "positioning": "Uniform Q4",
+      "positioning": "Q4 reference",
       "product": "M",
+      "reasoning_support": "supported",
       "recommended": true,
       "sha256": "f8e45572b9cc35161d4772b09bccfd383fe0bb03fc6d69b40a9138731302290b",
       "tensor_count": 426,
@@ -185,6 +200,7 @@
         "mmlu_pro": {
           "accuracy": 0.7958776595744681,
           "correct": 9576,
+          "not_reported_reason": null,
           "run_id": "release_20260715_quant_qwen35_mmlu_pro_cardlike_v2_qwen3p5_4b_l_mmlu_pro_official",
           "sample_count": 12032,
           "status": "complete"
@@ -200,8 +216,9 @@
       "lm_head_policy": "skip_tied",
       "output_embedding_mode": "tied_alias",
       "parameter_count": 4205751296,
-      "positioning": "Uniform Q8",
+      "positioning": "Maximum fidelity",
       "product": "L",
+      "reasoning_support": "supported",
       "recommended": false,
       "sha256": "b81d86c3ccbd1af44474597c0ef97af8e4d71ba40f3b5e7a20b4d6c2f709203c",
       "tensor_count": 426,