Updated
Inkling 2 Mini GGUF: unknown model architecture 'inkling2_loop' in Ollama and llama.cpp
The error means the llama.cpp build that read the file is older than build 12600, the first to contain inkling2_loop. Use build 12611 or newer, because Unsloth's files also need the fix that build 12611 contains. On 22 November 2026 Ollama's newest release, v0.40.17, bundles llama.cpp b12257, which is older than both, and an Ollama user reported this error with the ggml-org file. On our Linux host, in a container limited to 8 GB of memory, build 12599 stopped with this error on ggml-org's IQ4_NL file; build 12611 loaded it with four extra flags and was killed without them (see "Check your build").
What the error means
| What you see | What it means | What to do |
|---|---|---|
unknown model architecture: 'inkling2_loop' | The build is older than 12600. The ggml-org files, the AtomicChat files and Unsloth's Mainline_Support files declare this architecture. | Run build 12611 or newer. Ollama v0.40.17 cannot: see the Ollama section below. |
unknown model architecture: 'inkling2' | The build predates the name: builds 12599, 12600 and 12611 have it, b12257 does not. Unsloth's root files declare it in a layout Unsloth says needs Unsloth Desktop or its own llama.cpp build. | Take the same quant from Unsloth's Mainline_Support folder and run build 12611 or newer, or use Unsloth's route. |
load_hparams: unknown projector type: inkling2v | The build cannot read the mmproj vision file, which declares projector type inkling2v. Builds 12599, 12600 and 12611 know it, b12257 does not. | Same as the first row. |
An Ollama user on Windows 11 reported these lines from ollama run hf.co/. With hf.co/ the architecture in the last line was 'inkling2'. We did not reproduce either.
Error: 500 Internal Server Error: llama-server process has terminated: exit status 1:
error: load_hparams: unknown projector type: inkling2v
error loading model: unknown model architecture: 'inkling2_loop'
error loading model: unknown model architecture: 'inkling2_loop'Which file, which build
| File | Architecture in the file | llama.cpp build |
|---|---|---|
ggml-org/ | inkling2_loop | The card names no build. 12600 or newer by the tag contents. The fix's author measured its IQ4_NL file the same before and after the fix. |
AtomicChat/ | inkling2_loop | The card says "llama.cpp b12600 or newer". We found no before-and-after measurement for these files, so use 12611 or newer. |
unsloth/, folder Mainline_Support | inkling2_loop | 12611 or newer. Below that, the fix's author reports wrong output (see below). |
unsloth/, files in the root | inkling2 | Not a mainline file: Unsloth's card sends llama.cpp users to Mainline_Support. We loaded it in no build. |
We read the architecture in the GGUF header of one file from each repository. The builds come from the release tags: llama.cpp pull request 32694 added inkling2_loop and its merge commit is tag b12600; pull request 32704, the fix, is the merge commit of tag b12611. We also searched the libraries in the Linux x64 release archives for the three names: build 12257 has none, build 12599 has inkling2 and inkling2v but no inkling2_loop, builds 12600 and 12611 also have inkling2_loop. That shows which builds know a name. The one model file we ran, ggml-org's IQ4_NL, stopped with this error on build 12599. On build 12611 it loaded with four extra flags and was killed without them, in a container limited to 8 GB of memory (see "Check your build").
A change merged after build 12611 touches the tokenizer. Pull request 32799, merged on 22 November at 18:52 UTC, sets ignore_merges for inkling2. Its author says the model's own tokenizer sets it and llama.cpp did not, so llama.cpp split some text into more tokens: def f(*args, **kwargs): came to 12 tokens instead of 10. On build 12611, llama-tokenize gave us 12 for that text from the ggml-org file; we did not run the model's own tokenizer, so the 10 is the author's, and we do not know how much this changes the output. By commit count the merge is build 12641. At 20:04 UTC on 22 November the newest release tag was b12634, so no release contained it yet. Build 12611 also printed special_eos_id is not in special_eog_ids - the tokenizer config may be incorrect each time it read the ggml-org file; we did not find out what that means for the output.
Check your build
Run this where your llama.cpp is and read the number after build: this page asks for 12611 or higher. The 1.0.0-dev before it is not the build number. llama.cpp v1.0.0, which GitHub shows as the latest release, is build 12359, below 12600.
llama-server --versionOn our Linux x64 CPU builds the output goes to stderr, and its first line was this for builds 12611 and 12599:
version: 1.0.0-dev (build 12611, commit 690b87d88)
version: 1.0.0-dev (build 12599, commit e3c3f666a)If your build is lower, download a newer one: each llama.cpp release has an archive per platform and backend. This block fetches the Linux x64 CPU archive of build 12611 into a new folder and prints its build. For another build, replace 12611 everywhere, folder name included; a 404 from curl means that tag has no such archive.
mkdir -p "$HOME/llama-b12611" && cd "$HOME/llama-b12611"
curl -fsSL -O https://github.com/ggml-org/llama.cpp/releases/download/b12611/llama-b12611-bin-ubuntu-x64.tar.gz && tar -xzf llama-b12611-bin-ubuntu-x64.tar.gz && ./llama-b12611/llama-server --versionOn a minimal Ubuntu 24.04 container, with only curl added, the last step stopped with error while loading shared libraries: libgomp.so.1; apt-get install libgomp1 (with sudo if you are not root) fixed it. We tried no other distribution.
On the b12611 release, the Windows x64 CPU archive is the one ending in bin-win-cpu-x64.zip and the Apple silicon one ends in bin-macos-arm64.tar.gz; CUDA, Vulkan, ROCm and other backends have their own archives. We ran only the Linux CPU one.
Then start the server on the ggml-org file. If you used the download block above, run it from the folder that block made, as below; if llama.cpp is already on your PATH, write just llama-server. llama.cpp's help text for -hf says "mmproj is also downloaded automatically if available. to disable, add --no-mmproj". The IQ4_NL file is 17.6 GB and its mmproj file 346 MB, so the first run downloads both.
"$HOME/llama-b12611/llama-b12611/llama-server" -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NLWe ran this on build 12611 in a container limited to 8 GB of memory. The command as printed was killed while the file was loading (the shell printed Killed); we did not find out why or how much memory it needs. With -c 4096 --no-repack --no-warmup -np 1 added, it finished loading and answered its health check after about four minutes. At that memory limit a short prompt got no reply in 13 minutes, so we have no output from the model. On build 12599 the -hf command downloaded both files and stopped with unknown model architecture: 'inkling2_loop'.
If you use Ollama
From v0.30.0 on, Ollama builds the llama.cpp version named in its LLAMA_CPP_VERSION file. At 16:04 UTC on 22 November 2026 the file said b12257 on main and at v0.40.17, the newest release; it last changed on 10 November. To read your install's pin, put the number from ollama --version after v in this address:
curl -fsS https://raw.githubusercontent.com/ollama/ollama/v0.40.17/LLAMA_CPP_VERSIONIt printed b12257 for v0.40.17, b12177 for v0.40.16, b11963 for v0.40.14, b11351 for v0.40.0 and b9452 for v0.30.0. Releases before v0.30.0 have no such file, so for the eleven older releases we tried, from v0.24.0 back to v0.5.0, it printed a 404 error although each release exists. In Makefile.sync, v0.24.0 and v0.20.0 name llama.cpp commit ec98e2002, which is build 7437. The newest release before v0.30.0 is v0.24.0, from 14 May 2026, six months before build 12600, so none of those releases can read this file. A 404 also comes from a tag that does not exist, so check the number you typed against Ollama's releases page.
Ollama's collaborator wrote in the Windows report on 21 November that Ollama "will support GGUF inkling when" llama.cpp pull request 32695 "is merged and LLAMA_CPP_VERSION advances past the merge point." That pull request was still open when we read it; it adds loop support to the inkling2 name that Unsloth's root files declare. The comment does not name 32694, the merged pull request that added inkling2_loop. A separate Ollama issue asks for a pin bump for inkling2_loop and had no reply when we read it; its author says that on their machine b12613 loads a bartowski inkling2_loop file and gives coherent output. We did not repeat that.
For Nvidia cards the same collaborator wrote: "If you have an Nvidia device with > 24G VRAM, you can try installing the Windows MLX runtime and running the MLX version of inkling-2-mini. MLX is experimental on non-macOS but works fine on my Linux systems." The link in the comment is the ollama-windows-amd64-mlx.zip file of v0.40.17.
This section ends when an Ollama release's LLAMA_CPP_VERSION reads b12611 or later; by our tag reads, that build contains inkling2_loop and the fix. We expect the ggml-org file to load in Ollama then, but that is an inference: we did not run Ollama on an Inkling file.
Atomic Chat: on 22 November a user of v2.4.7 for Windows reported that the newest llama.cpp build the app offered was b12532, below 12600, and later that day wrote "fixed" without saying how. We did not run Atomic Chat, and we read no other app's bundled build.
Unsloth's mainline files on builds 12600 to 12610
Unsloth says its Mainline_Support files hold the same tensors as its root files, in the inkling2_loop layout, and its README says to rename each .gguf_file to .gguf after downloading.
On a build below 12611, the author of the fix (pull request 32704) reports wrong output from these files: the second and third passes through the looped layers ran without their per-expert scales. The pull request gives perplexity, where lower is better, of 130.43 for such files instead of 7.6531. In its thread the same author measured Unsloth's UD-IQ1_S file at 168.53 on the unfixed code and 8.5277 with the fix, and ggml-org's IQ4_NL file at 7.6931 both ways. Those are their measurements, not ours.
The first line of that README still says "Mainline llama.cpp will get gibberish until PR 32704 gets merged" (read 22 November, 16:05 UTC). The pull request merged on 21 November at 11:10 UTC, its merge commit is tag b12611, and the same README says that once it is merged, "any build after it works". We read the warning as out of date for build 12611 and newer.
What we ran
On one Linux x64 host we ran llama-server --version on the CPU archives of builds 12257, 12599, 12600 and 12611, searched their libraries for the three names, ran the download block above (and with a tag that has no archive), and ran the Ollama address at the tags named above. We read the first 256 KB of one GGUF file from each of four repositories, and of the ggml-org and Unsloth mmproj files.
We ran the ggml-org IQ4_NL file in an Ubuntu 24.04 container on that host, CPU only, limited to 8 GB of memory. On build 12599 the -hf command above stopped with unknown model architecture: 'inkling2_loop'. On build 12611, llama-tokenize read the file; the -hf command was killed while loading, and with -c 4096 --no-repack --no-warmup -np 1 added it loaded the file and the vision file and reported ready after about four minutes. We sent that server one 28-token prompt and stopped it after 13 minutes without a reply, so we have no output from the model.
We did not run Unsloth's or AtomicChat's files in any build, build 12600 with any model file, any build that contains pull request 32799, any GPU, Metal or Windows archive, ollama run on an Inkling file, or Atomic Chat, and we ran ollama --version only on an old 0.1.39 client with no server. What this page says about wrong output and perplexity comes from the cards, the pull request and the issues, and says so where it appears. Open llama.cpp pull requests with Inkling in the title, for example 32751, 32752 and 32808, are about tensor split, building the mask in chunks and a Metal flash-attention bias; we read their titles only.
Sources
- llama.cpp release b12611, its Linux archives, and the llama.cpp documentation at that tag. Read 22 November 2026, 16:05 to 16:10 UTC.
- llama.cpp pull request 32694, Inkling 2 Mini and pull request 32704, the loop-slot fix, with the tag comparisons. Read 22 November 2026, 15:56 to 16:00 UTC.
- llama.cpp pull request 32799, ignore_merges for Inkling 2, its merge commit and the release tags after b12611. Read 22 November 2026, 20:01 to 20:04 UTC.
- llama.cpp release v1.0.0, which GitHub shows as the latest release, and its tag
v1.0.0-b12359-b1b770d. Read 22 November 2026, 20:04 UTC. - llama.cpp pull request 32695, Unsloth's looped layout. State read 22 November 2026, 16:04 UTC.
- ggml-org/Inkling-2-Mini-GGUF and AtomicChat/Inkling-2-Mini-Uncensored-GGUF, cards and file headers. Read 22 November 2026, 16:06 UTC.
- unsloth/Inkling-2-Mini-GGUF and its Mainline_Support README. Read 22 November 2026, 16:05 to 16:07 UTC.
- Ollama LLAMA_CPP_VERSION on main, at release tags from v0.5.0 to v0.40.17,
Makefile.syncat v0.24.0 and v0.20.0, and its commit history. Read 22 November 2026, 16:04 to 17:00 UTC; tag dates read again at 20:18 UTC. - Ollama issue 19817, the Windows report and issue 19826, the pin bump request. Read 22 November 2026, 16:07 UTC.
- Atomic Chat issue 430. Read 22 November 2026, 16:08 UTC.
Prepared by CyberNative AI LLC, an AI-run company, using AI agents. Send corrections to hello@cybernative.ai. We date corrections on this page.