feat: advertise 262,144 native / 1,010,000 extensible context; bump default num_ctx

State the Qwen 3.6 context spec precisely instead of rounding to "~1 M":
262,144 native, extensible to 1,010,000 tokens with YaRN. The baked default
num_ctx moves 1000000 -> 1010000 in params and the Modelfile (bridge-synced),
superseding the earlier "match the Fable 1M budget" value. Retune the opt-in
YaRN recipe to --rope-scale 3.853 (1010000/262144) at -c 1010000 and recompute
the theoretical footprint at the default to ~62 GB KV / ~100 GB total. Native
(262,144) figures and all override examples are unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CHANGELOG.md

@@ -31,6 +31,16 @@ track the **tooling and documentation**, not the underlying base model.
   remains in the README "Recommended sampling" table as the tighter-reasoning
   override.
 
+### Changed (context-length ceiling raised to Qwen's 1,010,000 extensible max)
+- **Advertised context is now the precise 262,144-native / 1,010,000-extensible
+  pair, and the baked default rises to match.** The card previously rounded the
+  extensible window to `~1 M` and baked `num_ctx 1000000` (a Fable-teacher budget
+  match); both now use the exact Qwen 3.6 figures — **262,144 native, extensible
+  to 1,010,000 tokens with YaRN**. The baked default moves `1000000 → 1010000`
+  in `params` and the `Modelfile` (bridge-synced), the opt-in YaRN recipe is
+  retuned to `--rope-scale 3.853` (≈ `1010000 / 262144`) at `-c 1010000`, and the
+  theoretical footprint at the default is recomputed to ~62 GB KV / ~100 GB total.
+
 ### Fixed
 - **Hardware tables agree on the CPU-only RAM floor.** The README hardware
   table lists `≥48 GB RAM (CPU-only)` while the `Modelfile` working-configs

Modelfile

@@ -71,7 +71,7 @@ PARAMETER temperature 1.0
 PARAMETER top_p 0.95
 PARAMETER top_k 0
 PARAMETER repeat_penalty 1.05
-PARAMETER num_ctx 1000000
+PARAMETER num_ctx 1010000
 
 # Stop tokens. Without these, Ollama only honors <|im_end|> from the GGUF
 # metadata; the model occasionally emits <|endoftext|> instead and Ollama
@@ -99,23 +99,23 @@ Behavior rules:
 # This Q4_K_M is ~19 GB on disk. Real footprint at runtime:
 #   weights mmap          ~19 GB
 #   compute graph alloc   ~19 GB  (Ollama log: device.go:272 "total memory")
-#   KV cache @ 262K ctx   ~16 GB  (q8_0, ~2 GB / 32K; so ~61 GB at the 1M default)
-#   total minimum          ~99 GB  at the 1000000 default / ~53 GB at 262144 native
+#   KV cache @ 262K ctx   ~16 GB  (q8_0, ~2 GB / 32K; so ~62 GB at the 1.01M default)
+#   total minimum          ~100 GB at the 1010000 default / ~53 GB at 262144 native
 #                                  (theoretical, extrapolated from ~2 GB/32K; the
-#                                   default num_ctx is 1000000 — a ~1M ceiling ABOVE the
+#                                   default num_ctx is 1010000 — a ~1.01M ceiling ABOVE the
 #                                   262144 native context. No YaRN rope-scaling is
 #                                   baked in this GGUF, so output past ~262K degrades;
-#                                   treat 1M as an advertised ceiling. Most
+#                                   treat 1.01M as an advertised ceiling. Most
 #                                   hosts must override num_ctx down, see below)
 #
 # Working configurations (rows assume num_ctx trimmed to a practical ~16-32K;
-# at the 1M default the ~99 GB footprint fits only 128 GB+ hosts — override down
+# at the 1.01M default the ~100 GB footprint fits only 128 GB+ hosts — override down
 # everywhere else, see below):
 #   ✓ Single H100 80GB / A100 80GB              — full GPU offload
 #   ✓ RTX 5090 32GB / RTX 4090 24GB             — partial offload, ~15-25 tok/s
 #   ✓ Mac Studio M2/M3 Ultra 64GB+              — unified memory, ~20+ tok/s
 #   ✓ Linux box with 48GB+ RAM (CPU-only)       — ~3-6 tok/s
-#   ⚠ ASUS ROG Flow Z13 (Ryzen AI Max+, 32GB)   — OOMs at the 1M default (and
+#   ⚠ ASUS ROG Flow Z13 (Ryzen AI Max+, 32GB)   — OOMs at the 1.01M default (and
 #                                                 above ~32K); fits with num_ctx
 #                                                 ≤ 4096 and num_batch ≤ 256 (verified)
 #

README.md

@@ -89,7 +89,7 @@ ollama create janus -f Modelfile && ollama run janus
 
 After either path, `ollama show janus` lists `completion`, `tools`,
 and `thinking` under Capabilities. Hardware: the default `num_ctx` is
-**1000000** — a ~1M ceiling above the 262144 native window (YaRN is not baked
+**1010000** — a ~1.01M ceiling above the 262144 native window (YaRN is not baked
 into this GGUF, so context past ~262K degrades) — so trim it down to fit your
 host (see [Hardware requirements](#hardware-requirements)).
 
@@ -125,7 +125,7 @@ GGUF-only release. Pull the Heretic safetensors from [`llmfan46/Qwen3.6-35B-A3B-
 
 - Qwen 3.6, 35B total / 3B active, MoE (256 experts, 8 activated per token)
 - 40 layers, 10 × (3 × DeltaNet → MoE / 1 × Gated Attention → MoE)
-- 262 144 native context (extensible to ~1 M with YaRN, but YaRN is not enabled in the bundled GGUF)
+- 262 144 native context (extensible to 1 010 000 with YaRN, but YaRN is not enabled in the bundled GGUF)
 - Vision + video supported by upstream (mmproj not included in this release)
 - Vocab 248,320
 
@@ -295,7 +295,7 @@ needed for vision.
 
 ## Hardware requirements
 
-This is a ~19 GB Q4_K_M GGUF. Ollama's runtime footprint is **roughly 2× the model file** (weights mmap + compute graph), plus a KV cache that scales ~2 GB per 32K (q8_0). The default `num_ctx` is **1000000** — a ~1M ceiling above the 262144 native window — so KV alone is ~61 GB for ~99 GB total (theoretical, extrapolated from the ~2 GB/32K rule). This GGUF ships **no YaRN rope-scaling** (`rope.freq_base 10M`, no `rope.scaling`), so positions past the 262144 native window use untrained RoPE and output degrades — treat 1M as an advertised ceiling and keep real work within ~262K. **Most hosts must override `num_ctx` down**: e.g. the 262144 native window → ~16 GB KV / ~53 GB total, or `num_ctx 32768` → ~2 GB KV / ~39 GB total. 32 GB hosts fit the model by trimming ctx + batch (see Z13 row in the table).
+This is a ~19 GB Q4_K_M GGUF. Ollama's runtime footprint is **roughly 2× the model file** (weights mmap + compute graph), plus a KV cache that scales ~2 GB per 32K (q8_0). The default `num_ctx` is **1010000** — a ~1.01M ceiling above the 262144 native window — so KV alone is ~62 GB for ~100 GB total (theoretical, extrapolated from the ~2 GB/32K rule). This GGUF ships **no YaRN rope-scaling** (`rope.freq_base 10M`, no `rope.scaling`), so positions past the 262144 native window use untrained RoPE and output degrades — treat 1.01M as an advertised ceiling and keep real work within ~262K. **Most hosts must override `num_ctx` down**: e.g. the 262144 native window → ~16 GB KV / ~53 GB total, or `num_ctx 32768` → ~2 GB KV / ~39 GB total. 32 GB hosts fit the model by trimming ctx + batch (see Z13 row in the table).
 
 **How to override it:** `ollama run` has no `-o` flag, and `OLLAMA_CONTEXT_LENGTH`
 only sets a default that the baked `num_ctx` overrides — so set it per-session
@@ -317,16 +317,16 @@ Programmatic callers pass the same via the API `options` field:
 | Single H100/A100 80 GB | Works, full offload, ~30+ tok/s |
 | RTX 4090 24 GB / 5090 32 GB + 32 GB RAM | Works, partial offload, ~15-25 tok/s |
 | Mac Studio M2/M3 Ultra 64 GB+ unified | Works, ~20+ tok/s |
-| 32 GB unified-memory laptops (Ryzen AI Max+, Apple M-series) | Works with `num_ctx ≤ 4096` and `num_batch ≤ 256` to fit the compute graph; the 1M default OOMs (override `num_ctx` down). Measured 28.71 tok/s on ASUS ROG Flow Z13 GZ302EA at Q4_K_M (Radeon 8060S iGPU via ROCm gfx1151). |
+| 32 GB unified-memory laptops (Ryzen AI Max+, Apple M-series) | Works with `num_ctx ≤ 4096` and `num_batch ≤ 256` to fit the compute graph; the 1.01M default OOMs (override `num_ctx` down). Measured 28.71 tok/s on ASUS ROG Flow Z13 GZ302EA at Q4_K_M (Radeon 8060S iGPU via ROCm gfx1151). |
 
-**Reaching a coherent ~1M context (opt-in YaRN).** The bundled GGUF ships no YaRN rope-scaling, so the 1M default degrades past the 262144 native window (see above). Ollama has no rope knob, so for a genuinely coherent long context run the GGUF under llama.cpp with YaRN enabled:
+**Reaching a coherent ~1.01M context (opt-in YaRN).** The bundled GGUF ships no YaRN rope-scaling, so the 1.01M default degrades past the 262144 native window (see above). Ollama has no rope knob, so for a genuinely coherent long context run the GGUF under llama.cpp with YaRN enabled:
 
 ```bash
 llama-server -m Janus-35B-A3B.Q4_K_M.gguf \
-  --rope-scaling yarn --yarn-orig-ctx 262144 --rope-scale 3.8 -c 1000000
+  --rope-scaling yarn --yarn-orig-ctx 262144 --rope-scale 3.853 -c 1010000
 ```
 
-`--rope-scale 3.8` ≈ 1000000 / 262144; use a smaller factor for a smaller window. Static YaRN rescales *all* prompts, so enable it only when you actually need > 262K — it slightly degrades short-context quality otherwise.
+`--rope-scale 3.853` ≈ 1010000 / 262144; use a smaller factor for a smaller window. Static YaRN rescales *all* prompts, so enable it only when you actually need > 262K — it slightly degrades short-context quality otherwise.
 
 ## Chat template
 

params

@@ -3,7 +3,7 @@
   "top_p": 0.95,
   "top_k": 0,
   "repeat_penalty": 1.05,
-  "num_ctx": 1000000,
+  "num_ctx": 1010000,
   "stop": [
     "<|im_end|>",
     "<|endoftext|>",