Modelfile

# Thanatos-27B — Ollama wrapper around Qwen 3.6 27B (dense)
#
# Text + tool calling. Vision via Ollama is currently broken for this
# architecture (ollama/ollama#15898 — the qwen35 arch entries are in
# Ollama's Go text engine but missing from the C++ llama.cpp fallback
# Ollama uses when an mmproj is attached). Use llama.cpp directly for
# image input, or wait for the fix. See the Vision section in README.md.
#
# This repo bundles a single GGUF: Thanatos-27B.Q4_K_M.gguf (~17 GB),
# stamped `general.architecture: 'qwen35'` — the upstream-canonical
# arch entry every released llama.cpp / Ollama loads under for the
# Qwen 3.5 / 3.6 hybrid SSM + attention family. `ollama create
# thanatos-27b -f Modelfile && ollama run thanatos-27b` loads it
# directly. See README "Architecture" for the full stamp history
# (eight flips between qwen35 and qwen36, settled on qwen35 at
# `e03e10e` after the 4th qwen36 round trip had its friction
# re-tested in a fresh next-day session).
#
# For other quants (Q3_K_M, Q5_K_M, Q6_K, etc.), `make build QUANT=Q3_K_M`
# downloads the chosen quant from llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF
# (filename pattern Qwen3.6-27B-uncensored-heretic-v2-${QUANT}.gguf) and
# patches FROM in a temp Modelfile copy. Note: no Q3_K_S in this repo;
# use Q3_K_M for the smallest practical quant.
#
# Other GGUF sources (use with `make build GGUF_PATH=...`):
#     https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF        # primary (this repo's default)
#     https://huggingface.co/llmfan46/Qwen3.6-27B-uncensored-heretic-v2-Native-MTP-Preserved-GGUF  # MTP head preserved
#     https://huggingface.co/unsloth/Qwen3.6-27B-GGUF                               # vanilla Qwen 3.6 (pre-Heretic)
#     https://huggingface.co/rico03/Qwen3.6-27B-Claude-Opus-Reasoning-Distilled-GGUF

FROM ./Thanatos-27B.Q4_K_M.gguf

# Chat template — Qwen 3.6 ChatML in Ollama Go-template form, with the
# tool-calling blocks Ollama's capability detector looks for. Without a
# TEMPLATE that references .Tools and .ToolCalls, /api/chat and
# /v1/chat/completions reject any request carrying a `tools` array with
# `<model> does not support tools`. Same template as the 35B sibling —
# both share the Qwen 3.6 chat format.
TEMPLATE """{{- $lastUserIdx := -1 -}}
{{- range $idx, $msg := .Messages -}}
{{- if eq $msg.Role "user" }}{{ $lastUserIdx = $idx }}{{ end -}}
{{- end }}
{{- if or .System .Tools }}<|im_start|>system
{{ if .System }}{{ .System }}

{{ end }}
{{- if .Tools }}# Tools

You may call one or more functions to assist with the user query.

You are provided with function signatures within <tools></tools> XML tags:
<tools>
{{- range .Tools }}
{"type": "function", "function": {{ .Function }}}
{{- end }}
</tools>

For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call>
{{- end -}}<|im_end|>
{{ end }}
{{- range $i, $_ := .Messages }}
{{- $last := eq (len (slice $.Messages $i)) 1 -}}
{{- if eq .Role "user" }}<|im_start|>user
{{ .Content }}<|im_end|>
{{ else if eq .Role "assistant" }}<|im_start|>assistant
{{ if (and $.IsThinkSet (and .Thinking (or $last (gt $i $lastUserIdx)))) -}}
<think>{{ .Thinking }}</think>
{{ end -}}
{{ if .Content }}{{ .Content }}{{ end }}
{{- if .ToolCalls }}
{{- range .ToolCalls }}
<tool_call>
{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}
</tool_call>
{{- end }}
{{- end }}{{ if not $last }}<|im_end|>
{{ end }}
{{- else if eq .Role "tool" }}<|im_start|>user
<tool_response>
{{ .Content }}
</tool_response><|im_end|>
{{ end }}
{{- if and (ne .Role "assistant") $last }}<|im_start|>assistant
{{ if and $.IsThinkSet (not $.Think) -}}
<think>

</think>

{{ else -}}
<think>
{{ end -}}
{{ end }}
{{- end }}"""

# Sampling tuned for reasoning + general use. See README "Recommended sampling"
# for creative/RP alternatives.
PARAMETER temperature 1.0
PARAMETER top_p 0.95
PARAMETER top_k 0
PARAMETER repeat_penalty 1.05
PARAMETER num_ctx 1000000

# Stop tokens. Without these, Ollama only honors <|im_end|> from the GGUF
# metadata; the model occasionally emits <|endoftext|> instead and Ollama
# keeps generating past it (synthesising a fake new user turn). Listing
# both — plus <|im_start|> as a belt-and-braces guard against the same
# loop — keeps responses cleanly terminated.
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER stop "<|im_start|>"

SYSTEM """You are Thanatos, a precise and capable assistant for reasoning, writing, coding, and long-form dialogue.

Behavior rules:
- Answer the user's actual request directly.
- Be accurate, complete, and structured.
- Think before answering, but do not get stuck in repetitive loops or meta-commentary.
- If the request is ambiguous or incomplete, state what is missing and make the smallest reasonable assumption needed to continue.
- If the user wants creative writing, preserve tone, continuity, and character consistency.
- If the user wants analysis or technical help, prefer concrete steps, examples, and decisions over fluff.
- Finish with a usable answer, not just planning."""

# Hardware notes
# --------------
# Qwen 3.6 27B is *dense* — every parameter participates in every forward pass.
# Q4_K_M GGUF is ~17 GB. Practical footprint:
#   weights mmap          ~17 GB
#   compute graph alloc   ~12 GB  (smaller than 35B-A3B because dense ≠ MoE)
#   KV cache @ 262K ctx   ~16 GB  (q8_0, ~2 GB / 32K; so ~61 GB at the 1M default)
#   total minimum          ~90 GB  at the 1000000 default / ~45 GB at 262144 native
#                                  (theoretical, extrapolated from ~2 GB/32K; the
#                                   default num_ctx is 1000000 — a ~1M 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
#                                   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 ~90 GB footprint fits only 96 GB+ hosts — override down
# everywhere else, see below):
#   ✓ RTX 3090 / 4090 24 GB                     — full Q4 offload, ~25-40 tok/s
#   ✓ RTX 5090 32 GB                            — full offload at Q5/Q6 quant
#   ✓ Mac Studio M2/M3 32 GB+ unified           — ~15-25 tok/s
#   ✓ Linux box with 32 GB+ RAM (CPU-only)      — ~1-3 tok/s
#   ⚠ 32 GB unified-memory laptops              — borderline at Q4, try
#                                                 `make build QUANT=Q3_K_M`
#                                                 (~13 GB) and trim num_ctx
#
# Measured data points (ASUS ROG Flow Z13 GZ302EA, Ryzen AI Max+ 395 +
# Radeon 8060S iGPU, 32 GB unified, gfx1151, OLLAMA_FLASH_ATTENTION=1,
# OLLAMA_KV_CACHE_TYPE=q8_0, num_ctx 16384, 3-prompt mix):
#   Vulkan (OLLAMA_VULKAN=1):
#     Q3_K_S → 12.31 tok/s aggregate (run 1)
#       (6182 tokens / 501.9 s; 12.67 / 12.55 / 12.25 short/medium/long)
#     Q3_K_S → 11.70 tok/s aggregate (run 2, 2026-05-19 evening)
#       (8009 tokens / 684.0 s; 12.23 / 12.12 / 11.66 short/medium/long)
#       Second run measured against a `thanatos-27b:latest` (pre-rename)
#       built via `make build QUANT=Q3_K_S` against the then-current
#       unsloth/Qwen3.6-27B-GGUF source. Aggregate is 4.9% below
#       run 1 (within the ±20% noise band) — slightly longer
#       per-prompt outputs this run (8009 vs 6182 tokens) likely
#       contribute the difference, plus late-in-session thermal
#       pressure on the Strix Halo iGPU.
#       (Heretic v2 base is not benched here yet; rebundle pending.)
#     Q4_K_M →  9.31 tok/s aggregate (run 1)
#       (5356 tokens / 574.9 s;  9.48 /  9.43 /  9.28 short/medium/long)
#     Q4_K_M →  9.19 tok/s aggregate (run 2, 2026-05-19 afternoon)
#       (6210 tokens / 675.6 s;  9.40 /  9.29 /  9.16 short/medium/long)
#       Second run measured against the qwen36-stamped HF-bridge tag
#       after `make heal-hf` rebadged it to qwen35 in store — confirms
#       the in-place heal produces a model with the same performance
#       profile as `make load-bundle`. Aggregate is 1.3% below run 1
#       (within the ±20% noise band the README hardware section
#       warns about).
#     Q4_K_M →  9.32 tok/s aggregate (run 3, 2026-05-19 evening)
#       (4592 tokens / 492.7 s;  9.49 /  9.44 /  9.28 short/medium/long)
#       Third run, also against a heal-hf-rebadged qwen36-stamped
#       HF-bridge tag — this time the 3rd-round-trip bundle from
#       commit 973d7ef. Aggregate is within 0.1% of run 1's 9.31,
#       confirming the latest qwen36 -> qwen35 heal yields the same
#       performance profile as the prior two runs (no regression
#       from the third stamp flip).
#   ROCm (older snapshot, kept for backend comparison):
#     Q3_K_S → 10.14 tok/s aggregate
#       (8080 tokens / 796.5 s; 10.37 / 10.31 / 10.11 short/medium/long)