Add OpenClaw recipes and document VRAM-tiered context defaults by kenvandine · Pull Request #1 · lemonade-sdk/recipes

kenvandine · 2026-04-09T21:02:37Z

This PR adds a collection of model recipes curated for OpenClaw-style agent and assistant use cases.

Changes

Add openclaw/ directory with recipes for:
- GLM-4.7-Flash-GGUF
- Gemma-4-26B-A4B-GGUF and gemma-3-4b-it-GGUF (with corrected mmproj filenames)
- Qwen3-8B-GGUF (ctx_size bumped to 32768)
- Qwen3-Coder-30B-A3B-Instruct-GGUF
- Qwen3.5-9B-GGUF (vision + tool-calling)
- Qwen3.5-35B-A3B-Q4_K_M (vision + tool-calling)
- gpt-oss-20b-GGUF
Add openclaw/README.md documenting VRAM-tiered context size defaults
Update root README.md to reference the new openclaw/ directory

jeremyfowers · 2026-04-10T13:47:36Z

+    },
+    "model_name": "user.gemma-3-4b-it-GGUF",
+    "labels": [
+        "appear-builtin",


“appear-builtin” means “don’t show the user. in the name”, is that intentional here?

Main reason I ask is because the name gemma-3-4b-it-GGUF is very close to a builtin name.

Makes me wonder if we should have any special prefix or suffix for recipes. Like openclaw.gemma-3-4b-it-GGUF or gemma-3-4b-it-GGUF-OpenClaw

Good point. Switched to openclaw. prefix across all recipes and dropped appear-builtin — so the names will display as openclaw.gemma-3-4b-it-GGUF etc.

jeremyfowers · 2026-04-10T13:48:51Z

+        "main": "unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M",
+        "mmproj": "mmproj-F16.gguf"
+    },
+    "model_name": "user.Gemma-4-26B-A4B",


See comment above; in addition let’s make sure the naming scheme is consistent? e.g., the prior file had -GGUF but this one doesn’t.

Fixed — model_name is now openclaw.Gemma-4-26B-A4B-GGUF.

jeremyfowers · 2026-04-10T13:50:34Z

+    "recipe_options": {
+        "ctx_size": 32768
+    },
+    "size": 2.56


Crazy thought… if we’re hardcoding the context size on these, could we statically factor that into the size? That would make VRAM capacity based filtering much easier on the client side.

Let me think about that and come back to you.

jeremyfowers · 2026-04-10T13:51:43Z

@@ -0,0 +1,13 @@
+{
+    "checkpoint": "unsloth/Qwen2.5-Coder-7B-Instruct-128K-GGUF:Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf",


What’s the reason for using Qwen2.5 and Qwen3-VL? I would have thought the newer Qwen3.5 models of the same size would be strictly better.

It's just from the list of recommended models I found from other sources, now that I've proven the recipes mechanism works well with a decent collection of models I can look at refreshing that list and testing some of the newer models and remove some of the older models.

Checked — there's no Qwen3.5-Coder at ~7-9B (the coder line jumps to Qwen3-Coder-30B and Qwen3-Coder-Next at 80B). Replaced Qwen2.5-Coder-7B with Qwen3.5-9B (5.68 GB Q4_K_M), which is the better general model at the same size tier. For vision, Qwen2.5-VL-7B is dropped since Qwen3-VL-8B is already in the set and is the more capable vision-specific model.

All Qwen3.5 models are vision enabled and from my tests (I tested vision enabled model a lot) they are infinitely better than the older Qwen3-VL

Good to know — updated both Qwen3.5-9B and Qwen3.5-35B-A3B with mmproj and the vision label, and dropped the Qwen3-VL-8B and Qwen3-VL-30B recipes.

- Replace user. prefix with openclaw. and drop appear-builtin label on all recipes - Add -GGUF suffix to Gemma-4-26B-A4B model name for consistency - Replace Qwen2.5-Coder-7B with Qwen3.5-9B (newer generation, better across the board) - Remove Qwen2.5-VL-7B (superseded by Qwen3-VL-8B already in the set) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Qwen3.5 models are all vision-enabled and outperform Qwen3-VL. - Add mmproj and vision label to Qwen3.5-9B and Qwen3.5-35B-A3B - Remove Qwen3-VL-8B and Qwen3-VL-30B recipes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Adds --chat-template-kwargs to explicitly disable thinking mode, which may cause openclaw to not respond due to <think> blocks in the output. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Both models are small enough to fit in iGPU shared memory (2.3 GB and 4.1 GB at Q4_K_M) and support tool-calling via llama.cpp CPU/Vulkan backends. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Fix Phi-4-mini checkpoint to use quantization variant (Q4_K_M) instead of a full filename that doesn't exist in the repo. Remove Mistral-7B-v0.3 recipe as it requires HF authentication to download. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Drop the Q4_K_M variant since it can't find a matching file in the repo. Let lemonade pick the first .gguf file automatically. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Phi-4-mini was slow on Intel Vulkan and not responding in openclaw. Llama-3.2-3B has well-supported tool calling format and a similar size (~2 GB). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Fix reserveTokensFloor complaint by matching ctx_size used by other recipes. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

kenvandine added 5 commits April 9, 2026 11:31

Add OpenClaw recipes and document VRAM-tiered context defaults

c5f3be1

Fixed paths

216e475

Fix Gemma mmproj filenames and restore Qwen coding recipes

68f75f9

Raise context size for smaller OpenClaw vision models

28714f3

Bump ctx_size to 32768 for Qwen3-8B-GGUF

3fa6fb5

jeremyfowers assigned kenvandine Apr 10, 2026

jeremyfowers reviewed Apr 10, 2026

View reviewed changes

jeremyfowers requested review from bitgamma and sawansri April 10, 2026 13:52

kenvandine and others added 11 commits April 10, 2026 10:43

Revert to user. prefix pending server-side support for openclaw. prefix

35363cd

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Disable thinking mode for Gemma-4 openclaw recipe

8e505b9

Adds --chat-template-kwargs to explicitly disable thinking mode, which may cause openclaw to not respond due to <think> blocks in the output. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

openclaw: Gemma-4-26B-A4B isn't working with openclaw

94df9fd

openclaw: Add Phi-4-mini and Mistral-7B-v0.3 recipes for Intel iGPU

3701094

Both models are small enough to fit in iGPU shared memory (2.3 GB and 4.1 GB at Q4_K_M) and support tool-calling via llama.cpp CPU/Vulkan backends. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

openclaw: Remove Phi-4-mini checkpoint variant to use repo default

c51f066

Drop the Q4_K_M variant since it can't find a matching file in the repo. Let lemonade pick the first .gguf file automatically. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

openclaw: Replace Phi-4-mini with Llama-3.2-3B-Instruct

5cf1065

Phi-4-mini was slow on Intel Vulkan and not responding in openclaw. Llama-3.2-3B has well-supported tool calling format and a similar size (~2 GB). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

openclaw: Increase Llama-3.2-3B ctx_size to 32768

c50a366

Fix reserveTokensFloor complaint by matching ctx_size used by other recipes. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add OpenClaw recipe for Qwen3.6-35B-A3B-MTP-GGUF

cbe96f5

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add OpenClaw recipes and document VRAM-tiered context defaults#1

Add OpenClaw recipes and document VRAM-tiered context defaults#1
kenvandine wants to merge 16 commits into
lemonade-sdk:mainfrom
kenvandine:openclaw_recipes

kenvandine commented Apr 9, 2026 •

edited

Loading

Uh oh!

jeremyfowers Apr 10, 2026

Uh oh!

kenvandine Apr 10, 2026

Uh oh!

jeremyfowers Apr 10, 2026

Uh oh!

kenvandine Apr 10, 2026

Uh oh!

jeremyfowers Apr 10, 2026

Uh oh!

kenvandine Apr 10, 2026

Uh oh!

jeremyfowers Apr 10, 2026

Uh oh!

kenvandine Apr 10, 2026

Uh oh!

kenvandine Apr 10, 2026

Uh oh!

bitgamma Apr 10, 2026 •

edited

Loading

Uh oh!

kenvandine Apr 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		@@ -0,0 +1,13 @@
		{
		"checkpoint": "unsloth/Qwen2.5-Coder-7B-Instruct-128K-GGUF:Qwen2.5-Coder-7B-Instruct-Q4_K_M.gguf",

Conversation

kenvandine commented Apr 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

bitgamma Apr 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

kenvandine commented Apr 9, 2026 •

edited

Loading

bitgamma Apr 10, 2026 •

edited

Loading