Skip to content

AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the tools, so that you can focus on what matters.

feat(platform): recommend Ornith 1.5 for local users (#14188)

Nicholas Tindlec5609ef83dc0
diff --git a/autogpt_platform/backend/.env.default b/autogpt_platform/backend/.env.default
index c0f1192250..a090f9a099 100644
--- a/autogpt_platform/backend/.env.default
+++ b/autogpt_platform/backend/.env.default
@@ -126,8 +126,8 @@ CHAT_USE_LOCAL=false
#
# Under CHAT_USE_LOCAL=true these defaults are auto-rewritten to local-
# friendly Ollama slugs by GraphitiConfig._apply_local_graphiti_models:
-# GRAPHITI_LLM_MODEL=hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M (same as chat)
-# GRAPHITI_RERANKER_MODEL=hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
+# GRAPHITI_LLM_MODEL=hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M (same as chat)
+# GRAPHITI_RERANKER_MODEL=hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M
# GRAPHITI_EMBEDDER_MODEL=nomic-embed-text (needs ollama pull)
# Set GRAPHITI_*_MODEL to a non-cloud slug to bypass the auto-rewrite
# (your custom value passes through untouched). See
diff --git a/autogpt_platform/backend/backend/copilot/graphiti/config.py b/autogpt_platform/backend/backend/copilot/graphiti/config.py
index b0fab3e4dc..6b34c8f86f 100644
--- a/autogpt_platform/backend/backend/copilot/graphiti/config.py
+++ b/autogpt_platform/backend/backend/copilot/graphiti/config.py
@@ -25,13 +25,13 @@ _DEFAULT_RERANKER_MODEL = "gpt-4.1-nano"
_DEFAULT_EMBEDDER_MODEL = "text-embedding-3-small"
# Local-transport defaults. Mirrors dev's chat-side ``--with-ollama``
-# default (``hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M``, per
-# ``docs/platform/copilot-local-llm.md`` — 4B params, ~3.4GB resident,
-# 256k native context, vetted for OpenAI-shim structured output). The
+# default (``hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M``, per
+# ``docs/platform/copilot-local-llm.md`` — 9B params, ~5.8GB model file,
+# 262k native context, and OpenAI-compatible tool calling). The
# reranker reuses the same model since the prompts are simpler than
# extraction and pulling a second model for it would double the local
# install's disk + RAM footprint.
-_LOCAL_LLM_MODEL = "hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M"
+_LOCAL_LLM_MODEL = "hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M"
_LOCAL_RERANKER_MODEL = _LOCAL_LLM_MODEL
# Embedder defaults to Ollama's ``nomic-embed-text`` (~270MB, 768-dim
# vectors, well-supported by the OpenAI ``/v1/embeddings`` shim).
@@ -275,10 +275,10 @@ class GraphitiConfig(BaseSettings):
still at their cloud defaults and the operator hasn't pinned
a ``GRAPHITI_*_MODEL`` override:
- - ``llm_model`` (``gpt-4.1-mini``) → ``hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M``
+ - ``llm_model`` (``gpt-4.1-mini``) → ``hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M``
(matches dev's chat default — one Ollama pull powers both
surfaces; vetted for structured output / tool-calling shape).
- - ``reranker_model`` (``gpt-4.1-nano``) → same Qwen slug.
+ - ``reranker_model`` (``gpt-4.1-nano``) → same Ornith slug.
Reranker prompts are simpler than extraction; reusing the
chat model avoids pulling a second model just for reranking.
- ``embedder_model`` (``text-embedding-3-small``) →
diff --git a/autogpt_platform/backend/backend/copilot/graphiti/config_test.py b/autogpt_platform/backend/backend/copilot/graphiti/config_test.py
index 8c99f549a3..c33f082c87 100644
--- a/autogpt_platform/backend/backend/copilot/graphiti/config_test.py
+++ b/autogpt_platform/backend/backend/copilot/graphiti/config_test.py
@@ -193,11 +193,11 @@ class TestApplyLocalGraphitiModels:
) -> None:
_patch_chat_cfg(monkeypatch, _local_chat_cfg())
cfg = GraphitiConfig()
- # llm_model: gpt-4.1-mini → Qwen3.5-4B Unsloth GGUF
- assert cfg.llm_model == "hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M"
+ # llm_model: gpt-4.1-mini → Ornith-1.5-9B GGUF
+ assert cfg.llm_model == "hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M"
# reranker reuses the same model (simpler prompts, avoids a
# second Ollama pull).
- assert cfg.reranker_model == "hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M"
+ assert cfg.reranker_model == "hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M"
# embedder: text-embedding-3-small → nomic-embed-text
assert cfg.embedder_model == "nomic-embed-text"
diff --git a/autogpt_platform/backend/backend/copilot/local_context_probe.py b/autogpt_platform/backend/backend/copilot/local_context_probe.py
index 15aef47164..ff387ab45b 100644
--- a/autogpt_platform/backend/backend/copilot/local_context_probe.py
+++ b/autogpt_platform/backend/backend/copilot/local_context_probe.py
@@ -41,9 +41,9 @@ logger = logging.getLogger(__name__)
_FLOOR_RESERVE = 24_000
_TARGET_FLOOR = 4_096
-# Used when no backend reports a window. Matches the installer's default
-# ``OLLAMA_CONTEXT_LENGTH`` so the common Ollama path is correct even when the
-# model is briefly unloaded and ``/api/ps`` is empty.
+# Used when no backend reports a window. Deliberately remains 32k even though
+# the default Ornith installer uses 262k: assuming a huge window for a backend
+# that cannot report one risks overflowing its actual context limit.
LOCAL_CONTEXT_FALLBACK = 32_768
# Below this, the ~19k floor leaves almost no room for conversation — the
diff --git a/autogpt_platform/backend/backend/copilot/local_context_probe_test.py b/autogpt_platform/backend/backend/copilot/local_context_probe_test.py
index 622f6344fd..402e12aeb4 100644
--- a/autogpt_platform/backend/backend/copilot/local_context_probe_test.py
+++ b/autogpt_platform/backend/backend/copilot/local_context_probe_test.py
@@ -50,6 +50,9 @@ class TestCompactionTargetForWindow:
def test_64k_window(self):
assert compaction_target_for_window(65_536) == 41_536
+ def test_ornith_262k_window(self):
+ assert compaction_target_for_window(262_144) == 238_144
+
def test_tiny_window_clamps_to_floor(self):
assert compaction_target_for_window(4_096) == 4_096
assert compaction_target_for_window(1_000) == 4_096
diff --git a/autogpt_platform/installer/setup-autogpt.bat b/autogpt_platform/installer/setup-autogpt.bat
index 3298910f32..2ecc6a6377 100644
--- a/autogpt_platform/installer/setup-autogpt.bat
+++ b/autogpt_platform/installer/setup-autogpt.bat
@@ -11,7 +11,7 @@ REM /with-ollama Install Ollama (via winget), pull a default chat
REM model, and wire backend\.env so AutoPilot runs
REM without any cloud API keys (CHAT_USE_LOCAL=true).
REM See docs/platform/copilot-local-llm.md.
-REM /ollama-model=NAME Model to pull (default: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M).
+REM /ollama-model=NAME Model to pull (default: hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M).
REM /ollama-host=URL Use an existing Ollama at this URL instead of
REM installing one locally. Skips the Ollama install
REM but still writes the CHAT_USE_LOCAL .env entries.
@@ -24,7 +24,7 @@ set REPO_DIR=%SCRIPT_DIR%..\..
set CLONE_NEEDED=0
set LOG_FILE=
set WITH_OLLAMA=0
-set OLLAMA_MODEL=hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
+set OLLAMA_MODEL=hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M
set OLLAMA_HOST_URL=
REM --- Parse args ---
@@ -280,7 +280,7 @@ REM
REM - OLLAMA_HOST=0.0.0.0:11434 so containers can reach it via
REM host.docker.internal ^(which Docker Desktop on Windows auto-injects
REM in every container's /etc/hosts^).
-REM - OLLAMA_CONTEXT_LENGTH=32768 because the OpenAI shim does NOT
+REM - OLLAMA_CONTEXT_LENGTH=262144 because the OpenAI shim does NOT
REM honor options.num_ctx in the request body ^(ollama/ollama#2714^);
REM Ollama silently caps every request at the 4 k default otherwise,
REM truncating AutoPilot's ~8 k system prompt.
@@ -290,9 +290,9 @@ REM — so we also export into this shell so the readiness probe below
REM and any post-script user shells see the new values immediately.
echo Setting OLLAMA_HOST and OLLAMA_CONTEXT_LENGTH ^(user env^)...
setx OLLAMA_HOST "0.0.0.0:11434" >nul
-setx OLLAMA_CONTEXT_LENGTH "32768" >nul
+setx OLLAMA_CONTEXT_LENGTH "262144" >nul
set OLLAMA_HOST=0.0.0.0:11434
-set OLLAMA_CONTEXT_LENGTH=32768
+set OLLAMA_CONTEXT_LENGTH=262144
REM Restart Ollama so it picks up the new env. The Windows installer
REM registers Ollama as a background app that re-spawns on user login;
diff --git a/autogpt_platform/installer/setup-autogpt.sh b/autogpt_platform/installer/setup-autogpt.sh
index 1ce710ca4c..37cde84bf2 100755
--- a/autogpt_platform/installer/setup-autogpt.sh
+++ b/autogpt_platform/installer/setup-autogpt.sh
@@ -12,7 +12,7 @@
# wire backend/.env so AutoPilot runs without any
# cloud API keys (CHAT_USE_LOCAL=true). See
# docs/platform/copilot-local-llm.md.
-# --ollama-model=NAME Model to pull (default: hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M).
+# --ollama-model=NAME Model to pull (default: hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M).
# --ollama-host=URL Use an existing Ollama at this URL instead of
# installing one locally. Skips the Ollama install
# but still writes the CHAT_USE_LOCAL .env entries.
@@ -33,7 +33,7 @@ DOCKER_CMD="docker"
DOCKER_COMPOSE_CMD="docker compose"
LOG_FILE=""
WITH_OLLAMA=false
-OLLAMA_MODEL="hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M"
+OLLAMA_MODEL="hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M"
OLLAMA_HOST_URL=""
# OS-family detection — Linux and macOS install Ollama very differently
@@ -175,7 +175,7 @@ bootstrap_ollama() {
# the first chat turn fails with "model not found" — a setup
# script that returns 0 should mean the platform is *usable*.
# ``grep -F`` (fixed string) so model names containing regex
- # metacharacters (``hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M`` has ``.`` and
+ # metacharacters (``hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M`` has ``.`` and
# ``:``) match literally, not as wildcards.
if ! curl -sf "${OLLAMA_ROOT}/api/tags" \
| grep -Fq "\"name\":\"$OLLAMA_MODEL\""; then
@@ -234,7 +234,7 @@ _bootstrap_ollama_linux() {
sudo tee /etc/systemd/system/ollama.service.d/host.conf > /dev/null <<'OLLAMA_DROPIN'
[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"
-Environment="OLLAMA_CONTEXT_LENGTH=32768"
+Environment="OLLAMA_CONTEXT_LENGTH=262144"
OLLAMA_DROPIN
sudo systemctl daemon-reload
sudo systemctl restart ollama
@@ -288,7 +288,7 @@ _bootstrap_ollama_macos() {
# effect.
launchctl setenv OLLAMA_HOST "0.0.0.0:11434" \
|| handle_error "launchctl setenv OLLAMA_HOST failed"
- launchctl setenv OLLAMA_CONTEXT_LENGTH "32768" \
+ launchctl setenv OLLAMA_CONTEXT_LENGTH "262144" \
|| handle_error "launchctl setenv OLLAMA_CONTEXT_LENGTH failed"
# Stop any currently-running Ollama (tray app, brew service, raw
# ``ollama serve``). ``pkill`` matches by name; ``|| true`` so the
@@ -310,7 +310,7 @@ _bootstrap_ollama_macos() {
# Export the same vars on the serve invocation so the headless server
# honors them too. ``disown`` so the background job survives this script's
# exit even in shells with ``shopt -s huponexit`` (login shells, some CI).
- OLLAMA_HOST="0.0.0.0:11434" OLLAMA_CONTEXT_LENGTH="32768" \
+ OLLAMA_HOST="0.0.0.0:11434" OLLAMA_CONTEXT_LENGTH="262144" \
nohup ollama serve >/dev/null 2>&1 &
disown 2>/dev/null || true
for _ in $(seq 1 20); do
diff --git a/autogpt_platform/single-container/.env.example b/autogpt_platform/single-container/.env.example
index 7e4f855910..873624d8e7 100644
--- a/autogpt_platform/single-container/.env.example
+++ b/autogpt_platform/single-container/.env.example
@@ -56,7 +56,7 @@ CHAT_USE_OPENROUTER=true
CHAT_USE_LOCAL=false
CHAT_BASE_URL=
CHAT_API_KEY=
-# CHAT_FAST_STANDARD_MODEL=hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
+# CHAT_FAST_STANDARD_MODEL=hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M
# CHAT_FAST_ADVANCED_MODEL=qwen3:14b-q4_K_M
# CHAT_THINKING_STANDARD_MODEL=
# CHAT_THINKING_ADVANCED_MODEL=
diff --git a/docs/platform/copilot-local-llm.md b/docs/platform/copilot-local-llm.md
index 6a17d2f725..725b37ac0a 100644
--- a/docs/platform/copilot-local-llm.md
+++ b/docs/platform/copilot-local-llm.md
@@ -84,7 +84,7 @@ CHAT_API_KEY=ollama
# The chat model. Bare model names ONLY — provider/model slugs (e.g.
# `anthropic/claude-...`) are passed through verbatim and Ollama can't
# resolve them. See "Picking a model" below.
-CHAT_FAST_STANDARD_MODEL=hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
+CHAT_FAST_STANDARD_MODEL=hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M
# Optional — override for the advanced tier. If you leave it out, the
# local transport derives title_model, simulation_model, AND
@@ -104,14 +104,14 @@ model that handles all three.
| Tier | Recommended Ollama tag | Why | Footprint |
| --- | --- | --- | --- |
-| **Default** | `hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M` | Unsloth-recommended; solid OpenAI-shim tool-calling at 4B; 256 k native context; reasoning model (the chat UI renders its thinking separately from the answer) | ~3.4 GB resident; runs on a 16 GB laptop, GPU-accelerated |
+| **Default** | `hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M` | Official Ornith GGUF; agentic 9B model with OpenAI-compatible tool calling; 262,144-token native context; reasoning model (the chat UI renders its thinking separately from the answer) | ~5.8 GB model file; allow additional RAM for context and the KV cache |
| **Tight RAM** | `qwen3:4b` | Smaller; native tools; set `think: false` to avoid the unclosed-`<think>` tool-call render bug | ~3-4 GB resident |
| **GPU / advanced** | `qwen3:14b-q4_K_M` | Best tool-selection accuracy in this size class | ~12 GB VRAM |
Pull whichever you choose:
```bash
-ollama pull hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
+ollama pull hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M
```
## Context window — set it once, on the backend
@@ -129,8 +129,12 @@ backend's *actual* loaded window back at runtime — Ollama `/api/ps`, llama.cpp
it. Backends that don't report a window (LiteLLM proxy, Jan,
text-generation-webui) fall back to assuming 32k.
-Use **at least 24k** (32768 recommended): below that, the system prompt +
-tools leave almost no room for conversation and AutoPilot logs a warning.
+The default Ornith model has a 262,144-token native window, so the installer
+sets `OLLAMA_CONTEXT_LENGTH=262144`. This maximizes available conversation
+history but substantially increases KV-cache RAM/VRAM use. Operators using a
+custom model or constrained hardware can lower it, but should keep at least
+24k: below that, the system prompt + tools leave almost no room for
+conversation and AutoPilot logs a warning.
The installer sets `OLLAMA_CONTEXT_LENGTH` for you. Manual setup per platform:
@@ -140,7 +144,7 @@ The installer sets `OLLAMA_CONTEXT_LENGTH` for you. Manual setup per platform:
# /etc/systemd/system/ollama.service.d/host.conf
[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"
-Environment="OLLAMA_CONTEXT_LENGTH=32768"
+Environment="OLLAMA_CONTEXT_LENGTH=262144"
```
Then `sudo systemctl daemon-reload && sudo systemctl restart ollama`.
@@ -149,7 +153,7 @@ Then `sudo systemctl daemon-reload && sudo systemctl restart ollama`.
```bash
launchctl setenv OLLAMA_HOST 0.0.0.0:11434
-launchctl setenv OLLAMA_CONTEXT_LENGTH 32768
+launchctl setenv OLLAMA_CONTEXT_LENGTH 262144
# Then restart Ollama — either:
brew services restart ollama # if installed via the brew formula
# …or quit the menu-bar app and relaunch it (the .dmg install)
@@ -159,13 +163,13 @@ brew services restart ollama # if installed via the brew formula
```powershell
setx OLLAMA_HOST "0.0.0.0:11434"
-setx OLLAMA_CONTEXT_LENGTH "32768"
+setx OLLAMA_CONTEXT_LENGTH "262144"
# Then quit Ollama from the system tray and relaunch it
# (setx writes to HKCU but does NOT update already-running processes).
```
Verify on any platform with `ollama ps` (the `CONTEXT` column should
-show your value, e.g. 32768). If you change it, AutoPilot picks up the
+show your value, e.g. 262144). If you change it, AutoPilot picks up the
new window automatically on the next turn — nothing else to update.
## Networking — same host, different host, or remote
@@ -385,11 +389,11 @@ rewrites the cloud OpenAI defaults to local Ollama equivalents:
| Setting | Cloud default | Local default |
|---|---|---|
-| `GRAPHITI_LLM_MODEL` | `gpt-4.1-mini` | `hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M` |
-| `GRAPHITI_RERANKER_MODEL` | `gpt-4.1-nano` | `hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M` |
+| `GRAPHITI_LLM_MODEL` | `gpt-4.1-mini` | `hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M` |
+| `GRAPHITI_RERANKER_MODEL` | `gpt-4.1-nano` | `hf.co/ornith-ai/Ornith-1.5-9B-GGUF:Q4_K_M` |
| `GRAPHITI_EMBEDDER_MODEL` | `text-embedding-3-small` | `nomic-embed-text` |
-The LLM + reranker reuse the same Qwen 3.5 4B model the `--with-ollama`
+The LLM + reranker reuse the same Ornith 1.5 9B model the `--with-ollama`
installer already pulls for chat, so no extra `ollama pull` is needed
unless you've overridden them. The embedder is a separate model — see
the next section.