Drop the wash, split quant from cost, restore Context

Three fixes to the per-machine table:

No background on the cells. The Q6-or-better signal now rides on the
quant name alone, ink against muted, so the tier still reads without a
shaded ground. FITS_BG and the .swatch rule go with it, both unused after.

The quant and the memory it costs were one run of text ("Q8_0 30"), which
made them hard to tell apart. They are two spans now: the quant in mono
and bold, because it is a filename, and the cost muted, smaller and
carrying its unit beside it.

Context comes back. It was dropped to get the budget columns onto a phone
screen, which was not mine to decide. It sits after the budget columns
instead of before, so the answer still leads on a 390px screen.

The audit reads the two spans separately rather than splitting cell text
on a space, so a cell that loses its size fails instead of reparsing the
quant name as one. Both that and a doctored size were confirmed to fail.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: barretodavid

bench/audit.py

@@ -407,19 +407,31 @@ def chart_labels(svg):
         bad_cell.append(f"{label}: not a leaderboard row")
         continue
     files = (_mm.get(spec.partition("@")[0], {}).get("gguf") or {}).get("files") or {}
-    # the budget columns are the trailing cells, in VRAM_BUDGETS order
-    tds = re.findall(r'<td class="r[^"]*"[^>]*>(.*?)</td>', rest, re.S)[-len(VRAM_BUDGETS):]
+    # SCI leads and Context trails, so the budget columns are the cells between,
+    # in VRAM_BUDGETS order. Sliced from both ends rather than counted from one,
+    # so adding a trailing column cannot silently shift which cells are read.
+    tds = re.findall(r'<td class="r[^"]*"[^>]*>(.*?)</td>', rest, re.S)[1:-1]
+    if len(tds) != len(VRAM_BUDGETS):
+        bad_cell.append(f"{label}: {len(tds)} budget cells, expected {len(VRAM_BUDGETS)}")
+        continue
     for budget, td in zip(VRAM_BUDGETS, tds):
         want = best_quant_for(spec, budget)
-        got = re.sub(r"<[^>]+>", "", td).strip()
-        if not got:
+        # the quant and its cost are separate spans now; read them separately so a
+        # cell that loses one of them fails rather than reparsing as the other
+        qm = re.search(r'<span class="q">(.*?)</span>', td, re.S)
+        gm = re.search(r'<span class="gb">([\d,]+) GB</span>', td, re.S)
+        if not qm:
             if want:
                 bad_cell.append(f"{label} at {budget} GB: blank, expected {want[0]}")
             continue
         n_cell += 1
-        name, _, size = got.rpartition(" ")
-        if not want or name != want[0] or round(want[1]) != int(size.replace(",", "")):
-            bad_cell.append(f"{label} at {budget} GB: prints {got!r}, picker says {want}")
+        if not gm:
+            bad_cell.append(f"{label} at {budget} GB: names a quant with no size")
+            continue
+        name, size = qm.group(1).strip(), int(gm.group(1).replace(",", ""))
+        if not want or name != want[0] or round(want[1]) != size:
+            bad_cell.append(f"{label} at {budget} GB: prints {name} {size}, "
+                            f"picker says {want}")
         elif name not in files:
             bad_cell.append(f"{label} at {budget} GB: {name} is not in the published ladder")
         elif files[name] > budget - BUDGET_RESERVE[budget]:

bench/html_report.py

@@ -465,7 +465,6 @@ def head_to_head_chart(metrics, w=760):
 
 ATTEMPT_COLORS = [SNF_BLUE, "#7c7ba2", "#bab7df", "#cecde7"]  # 1, 2, 3, 4+ submissions
 UNSOLVED_COLOR = "#bdb5ad"  # never solved: the band that tops every column
-FITS_BG = "#dbe7fc"         # blue-05 wash: a quant that fits one local machine
 # Warm neutral on purpose. It has to be legible in a band 2% tall (DeepSeek
 # fails 1 run in 52), which ruled out the diagonal hatch tried first: at that
 # height it was invisible. It also has to stay off the lavender ramp, so it
@@ -1157,19 +1156,28 @@ def h2h_stats(spec):
                 cells += '<td class="r"></td>'
                 continue
             name, gb = pick
-            # Two channels rather than lightness alone, as before, but pointed at
-            # what varies now: every filled cell fits by construction, so the wash
-            # marks the QUALITY tier instead. Q6 and up keeps the waterline
-            # readable without comparing any two numbers.
+            # The file name and the memory it costs are two different kinds of
+            # thing, so they are set as two: the quant in mono, which is what it
+            # is, a filename, and the cost muted and smaller beside it. Reading
+            # one without the other was the complaint about "Q8_0 30".
+            #
+            # Tier still shows, in ink versus muted on the name alone. Q6 and up
+            # is where quantization stops being what holds the model back.
             tier = "fits" if gb * 8e9 / param_count(mm["params_total"]) >= 5.9 else "nofit"
             cells += (f'<td class="r {tier}" data-s="{gb:.1f}">'
-                      f'<b>{name}</b> {gb:,.0f}</td>')
+                      f'<span class="q">{name}</span> '
+                      f'<span class="gb">{gb:,.0f} GB</span></td>')
         open_rows_html.append(
             f'<tr><td>{r["label"]}</td>'
             # the index rides along so this table can be read in score order too,
             # which is the order it ships in; the models table above has the rest
             + num_td(r["sci"], f'{r["sci"]:.1f}')
-            + cells + "</tr>"
+            # Context sits after the budget columns rather than before them: it is
+            # reference, not the answer, and in front it pushed every budget column
+            # off a 390px screen.
+            + cells
+            + num_td(mm["context_length"], fmt_ctx(mm["context_length"]))
+            + "</tr>"
         )
 
     # every sortable table on the page, not one by id: the open-weight table
@@ -1224,10 +1232,10 @@ def h2h_stats(spec):
 <section>
   <h2>Which quantization for my machine</h2>
   <p class="takeaway" style="margin:0 0 10px"><b>Pick your memory, read across.</b> Each cell is the <b>best quantization that fits</b>, named as the file you would download from the {gguf_sources} GGUF repos, and what it weighs in GB. Every figure is a real published file, not arithmetic.</p>
-  <p class="takeaway" style="margin:0 0 10px">The weights do not get the whole machine, so each column reserves room for the OS, the KV cache and activations: {budget_reserve_note}. <span class="swatch" style="background:{FITS_BG}"></span> marks a pick at <b>Q6 or better</b>, where quantization is not the thing holding the model back.</p>
+  <p class="takeaway" style="margin:0 0 10px">The weights do not get the whole machine, so each column reserves room for the OS, the KV cache and activations: {budget_reserve_note}. A quant in <b>full black</b> is <b>Q6 or better</b>, where quantization is not the thing holding the model back; a greyed one is a 4-to-5-bit compromise.</p>
   <p class="takeaway" style="margin:0 0 10px">A blank means <b>nothing 4-bit or better fits</b>, which is a real answer rather than a missing one: below 4-bit degradation stops being minor and coding feels it first, so no cell recommends going lower. {too_big_note}</p>
   <div class="tablewrap"><table id="opentable" class="sortable">
-    <tr><th>Model</th><th class="r desc" data-num aria-sort="descending">SCI</th>{"".join(f'<th class="r" data-num>{b} GB</th>' for b in VRAM_BUDGETS)}</tr>
+    <tr><th>Model</th><th class="r desc" data-num aria-sort="descending">SCI</th>{"".join(f'<th class="r" data-num>{b} GB</th>' for b in VRAM_BUDGETS)}<th class="r" data-num>Context</th></tr>
     {"".join(open_rows_html)}
   </table></div>
 </section>"""
@@ -1331,10 +1339,15 @@ def h2h_stats(spec):
   table.sortable th:hover,table.sortable th:focus-visible{{color:var(--ink)}}
   table.sortable th.asc::after{{content:" \\25B2";font-size:9px}}
   table.sortable th.desc::after{{content:" \\25BC";font-size:9px}}
-  /* Q6-or-better vs a 4-to-5-bit compromise: shaded ground plus ink, against no
-     ground and muted text. Colour alone was too close to read at this size. */
-  #opentable td.fits{{background:{FITS_BG};color:var(--ink);font-weight:600}}
-  #opentable td.nofit{{color:var(--muted)}}
+  /* Two kinds of thing in one cell, set as two: the quant is a filename, so it
+     is mono and carries the ink, and the memory it costs is muted and smaller
+     beside it. No background wash on either. */
+  #opentable td.r .q{{font-family:var(--mono);font-weight:700;font-size:12.5px}}
+  #opentable td.r .gb{{color:var(--muted);font-size:11px;white-space:nowrap}}
+  #opentable td.fits .q{{color:var(--ink)}}
+  /* a 4-to-5-bit pick: still the best that fits, but quantization is now part of
+     what holds the model back, so the name drops to muted */
+  #opentable td.nofit .q{{color:var(--muted)}}
   /* Six narrow columns, not ten, so this one does not need the 900px floor the
      wide tables do. Lowering it is what puts the ANSWER columns on screen at
      390px: at 900 the phone showed Model and SCI and scrolled every budget
@@ -1344,7 +1357,6 @@ def h2h_stats(spec):
      across two lines and left the rows at uneven heights; the table scrolls, so
      there is no reason to break the cell instead. */
   #opentable td.r{{white-space:nowrap}}
-  .swatch{{display:inline-block;width:22px;height:12px;border-radius:2px;vertical-align:-1px}}
   td{{padding:7px 8px;border-bottom:1px solid var(--line);font-size:13px;vertical-align:middle}}
   th.r,td.r{{text-align:right}}
   .ci{{color:var(--muted);font-size:11px}}

results/report.html

@@ -57,10 +57,15 @@
   table.sortable th:hover,table.sortable th:focus-visible{color:var(--ink)}
   table.sortable th.asc::after{content:" \25B2";font-size:9px}
   table.sortable th.desc::after{content:" \25BC";font-size:9px}
-  /* Q6-or-better vs a 4-to-5-bit compromise: shaded ground plus ink, against no
-     ground and muted text. Colour alone was too close to read at this size. */
-  #opentable td.fits{background:#dbe7fc;color:var(--ink);font-weight:600}
-  #opentable td.nofit{color:var(--muted)}
+  /* Two kinds of thing in one cell, set as two: the quant is a filename, so it
+     is mono and carries the ink, and the memory it costs is muted and smaller
+     beside it. No background wash on either. */
+  #opentable td.r .q{font-family:var(--mono);font-weight:700;font-size:12.5px}
+  #opentable td.r .gb{color:var(--muted);font-size:11px;white-space:nowrap}
+  #opentable td.fits .q{color:var(--ink)}
+  /* a 4-to-5-bit pick: still the best that fits, but quantization is now part of
+     what holds the model back, so the name drops to muted */
+  #opentable td.nofit .q{color:var(--muted)}
   /* Six narrow columns, not ten, so this one does not need the 900px floor the
      wide tables do. Lowering it is what puts the ANSWER columns on screen at
      390px: at 900 the phone showed Model and SCI and scrolled every budget
@@ -70,7 +75,6 @@
      across two lines and left the rows at uneven heights; the table scrolls, so
      there is no reason to break the cell instead. */
   #opentable td.r{white-space:nowrap}
-  .swatch{display:inline-block;width:22px;height:12px;border-radius:2px;vertical-align:-1px}
   td{padding:7px 8px;border-bottom:1px solid var(--line);font-size:13px;vertical-align:middle}
   th.r,td.r{text-align:right}
   .ci{color:var(--muted);font-size:11px}
@@ -365,11 +369,11 @@ <h2 id="the-models"><a class="anchor" href="#the-models">The models</a></h2>
 <section>
   <h2 id="which-quantization-for-my-machine"><a class="anchor" href="#which-quantization-for-my-machine">Which quantization for my machine</a></h2>
   <p class="takeaway" style="margin:0 0 10px"><b>Pick your memory, read across.</b> Each cell is the <b>best quantization that fits</b>, named as the file you would download from the <a href="https://huggingface.co/bartowski">bartowski</a> and <a href="https://huggingface.co/unsloth">unsloth</a> GGUF repos, and what it weighs in GB. Every figure is a real published file, not arithmetic.</p>
-  <p class="takeaway" style="margin:0 0 10px">The weights do not get the whole machine, so each column reserves room for the OS, the KV cache and activations: 16 GB at 128, 8 GB at 64, 4 GB at 32, 4 GB at 24. <span class="swatch" style="background:#dbe7fc"></span> marks a pick at <b>Q6 or better</b>, where quantization is not the thing holding the model back.</p>
+  <p class="takeaway" style="margin:0 0 10px">The weights do not get the whole machine, so each column reserves room for the OS, the KV cache and activations: 16 GB at 128, 8 GB at 64, 4 GB at 32, 4 GB at 24. A quant in <b>full black</b> is <b>Q6 or better</b>, where quantization is not the thing holding the model back; a greyed one is a 4-to-5-bit compromise.</p>
   <p class="takeaway" style="margin:0 0 10px">A blank means <b>nothing 4-bit or better fits</b>, which is a real answer rather than a missing one: below 4-bit degradation stops being minor and coding feels it first, so no cell recommends going lower. The nine open models not listed need more than 128 GB even at 4-bit: Kimi K3, MiMo-V2.5-Pro, MiniMax M3, DeepSeek V4 Flash, Qwen3.8 Max, DeepSeek V4-Pro, Hy3, Inkling and GLM 5.2.</p>
   <div class="tablewrap"><table id="opentable" class="sortable">
-    <tr><th>Model</th><th class="r desc" data-num aria-sort="descending">SCI</th><th class="r" data-num>128 GB</th><th class="r" data-num>64 GB</th><th class="r" data-num>32 GB</th><th class="r" data-num>24 GB</th></tr>
-    <tr><td>Muse Glimmer 30B</td><td class="r" data-s="53.54335822195509">53.5</td><td class="r fits" data-s="29.6"><b>Q8_0</b> 30</td><td class="r fits" data-s="29.6"><b>Q8_0</b> 30</td><td class="r fits" data-s="24.1"><b>Q6_K_L</b> 24</td><td class="r nofit" data-s="19.4"><b>Q5_K_S</b> 19</td></tr><tr><td>gpt-oss-120b</td><td class="r" data-s="17.741231416968965">17.7</td><td class="r nofit" data-s="62.8"><b>Q4_K_M</b> 63</td><td class="r"></td><td class="r"></td><td class="r"></td></tr><tr><td>Qwen3.6-35B-A3B</td><td class="r" data-s="16.19082483091965">16.2</td><td class="r fits" data-s="36.9"><b>Q8_0</b> 37</td><td class="r fits" data-s="36.9"><b>Q8_0</b> 37</td><td class="r fits" data-s="26.6"><b>UD-Q5_K_XL</b> 27</td><td class="r nofit" data-s="19.5"><b>UD-IQ4_NL_XL</b> 20</td></tr><tr><td>Gemma 4 31B</td><td class="r" data-s="15.806948103252044">15.8</td><td class="r fits" data-s="32.6"><b>Q8_0</b> 33</td><td class="r fits" data-s="32.6"><b>Q8_0</b> 33</td><td class="r fits" data-s="27.5"><b>UD-Q6_K_XL</b> 28</td><td class="r nofit" data-s="18.3"><b>Q4_K_M</b> 18</td></tr><tr><td>Qwen3 Coder Next</td><td class="r" data-s="14.283251268776736">14.3</td><td class="r fits" data-s="84.8"><b>Q8_0</b> 85</td><td class="r nofit" data-s="55.8"><b>UD-Q5_K_S</b> 56</td><td class="r"></td><td class="r"></td></tr><tr><td>Qwen3.6-27B</td><td class="r" data-s="13.317904173386546">13.3</td><td class="r fits" data-s="28.6"><b>Q8_0</b> 29</td><td class="r fits" data-s="28.6"><b>Q8_0</b> 29</td><td class="r fits" data-s="25.6"><b>UD-Q6_K_XL</b> 26</td><td class="r fits" data-s="20.0"><b>UD-Q5_K_XL</b> 20</td></tr>
+    <tr><th>Model</th><th class="r desc" data-num aria-sort="descending">SCI</th><th class="r" data-num>128 GB</th><th class="r" data-num>64 GB</th><th class="r" data-num>32 GB</th><th class="r" data-num>24 GB</th><th class="r" data-num>Context</th></tr>
+    <tr><td>Muse Glimmer 30B</td><td class="r" data-s="53.54335822195509">53.5</td><td class="r fits" data-s="29.6"><span class="q">Q8_0</span> <span class="gb">30 GB</span></td><td class="r fits" data-s="29.6"><span class="q">Q8_0</span> <span class="gb">30 GB</span></td><td class="r fits" data-s="24.1"><span class="q">Q6_K_L</span> <span class="gb">24 GB</span></td><td class="r nofit" data-s="19.4"><span class="q">Q5_K_S</span> <span class="gb">19 GB</span></td><td class="r" data-s="131072">131k</td></tr><tr><td>gpt-oss-120b</td><td class="r" data-s="17.741231416968965">17.7</td><td class="r nofit" data-s="62.8"><span class="q">Q4_K_M</span> <span class="gb">63 GB</span></td><td class="r"></td><td class="r"></td><td class="r"></td><td class="r" data-s="131072">131k</td></tr><tr><td>Qwen3.6-35B-A3B</td><td class="r" data-s="16.19082483091965">16.2</td><td class="r fits" data-s="36.9"><span class="q">Q8_0</span> <span class="gb">37 GB</span></td><td class="r fits" data-s="36.9"><span class="q">Q8_0</span> <span class="gb">37 GB</span></td><td class="r fits" data-s="26.6"><span class="q">UD-Q5_K_XL</span> <span class="gb">27 GB</span></td><td class="r nofit" data-s="19.5"><span class="q">UD-IQ4_NL_XL</span> <span class="gb">20 GB</span></td><td class="r" data-s="262144">262k</td></tr><tr><td>Gemma 4 31B</td><td class="r" data-s="15.806948103252044">15.8</td><td class="r fits" data-s="32.6"><span class="q">Q8_0</span> <span class="gb">33 GB</span></td><td class="r fits" data-s="32.6"><span class="q">Q8_0</span> <span class="gb">33 GB</span></td><td class="r fits" data-s="27.5"><span class="q">UD-Q6_K_XL</span> <span class="gb">28 GB</span></td><td class="r nofit" data-s="18.3"><span class="q">Q4_K_M</span> <span class="gb">18 GB</span></td><td class="r" data-s="262144">262k</td></tr><tr><td>Qwen3 Coder Next</td><td class="r" data-s="14.283251268776736">14.3</td><td class="r fits" data-s="84.8"><span class="q">Q8_0</span> <span class="gb">85 GB</span></td><td class="r nofit" data-s="55.8"><span class="q">UD-Q5_K_S</span> <span class="gb">56 GB</span></td><td class="r"></td><td class="r"></td><td class="r" data-s="262144">262k</td></tr><tr><td>Qwen3.6-27B</td><td class="r" data-s="13.317904173386546">13.3</td><td class="r fits" data-s="28.6"><span class="q">Q8_0</span> <span class="gb">29 GB</span></td><td class="r fits" data-s="28.6"><span class="q">Q8_0</span> <span class="gb">29 GB</span></td><td class="r fits" data-s="25.6"><span class="q">UD-Q6_K_XL</span> <span class="gb">26 GB</span></td><td class="r fits" data-s="20.0"><span class="q">UD-Q5_K_XL</span> <span class="gb">20 GB</span></td><td class="r" data-s="262144">262k</td></tr>
   </table></div>
 </section>