Repository navigation
model : add LiquidAI/LFM3-1.7B and LiquidAI/LFM3-VL-2B - #31252
Merged
Merged
Conversation
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Xuan-Son Nguyen <son@huggingface.co> Assisted-by: Claude Opus 5.5
Assisted-by: Claude Opus 5.5
Assisted-by: Claude Opus 5.5
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Tarek Dakhran <github@dakhran.com>
Assisted-by: Claude Opus 5.5
Assisted-by: Claude Opus 5.5
tdakhran
requested review from
a team,
CISC,
ggerganov and
pwilkin
as code owners
October 27, 2026 15:05
github-actions
Bot
added
model
CISC
added
the
highlight
CISC
approved these changes
Oct 27, 2026
tdakhran
marked this pull request as draft
October 27, 2026 15:17
ngxson
approved these changes
Oct 27, 2026
Contributor
Author
|
A few final checks, uploading GGUFs, and regenerating the Metal Fusion CSV. |
Assisted-by: Claude Opus 5.5
Assisted-by: Claude Opus 5.5
tdakhran
marked this pull request as ready for review
October 27, 2026 15:36
7 of 10 tasks
This was referenced Oct 27, 2026
6 of 10 tasks
Closed
5 tasks
auRose94
pushed a commit
to auRose94/llama.cpp
that referenced
this pull request
Oct 28, 2026
* model : add LiquidAI/LFM3-1.7B and LiquidAI/LFM3-VL-2B Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Xuan-Son Nguyen <son@huggingface.co> Assisted-by: Claude Opus 5.5 * tests : set the SWA rope base of the LFM3 test model Assisted-by: Claude Opus 5.5 * llama : enable recurrent state rollback for LFM3 Assisted-by: Claude Opus 5.5 * Apply batched suggestions from code review Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Tarek Dakhran <github@dakhran.com> * gguf-py : align the lfm3vl mapping comments Assisted-by: Claude Opus 5.5 * llama : fix CRLF line endings from the review suggestions Assisted-by: Claude Opus 5.5 * tests : add the LFM3 Metal fusion baseline Assisted-by: Claude Opus 5.5 * llama : keep the LFM3-VL depthwise conv kernels in F32 when quantizing Assisted-by: Claude Opus 5.5 --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Xuan-Son Nguyen <son@huggingface.co>
timmyy123
added a commit
to timmyy123/LLM-Hub
that referenced
this pull request
Oct 28, 2026
Adds LiquidAI's LFM3 1.7B (262K context, reasoning, tool calling) and LFM3 VL 2B (images; llama.cpp has no LFM3-VL video yet) to the catalog. LFM3 needs llama.cpp b11946 (ggml-org/llama.cpp#31252), but an installed build was never updated, so existing installs would stay on one that can't load it. Windows, Linux and macOS now treat b11946 as the minimum: an older install is replaced on the next model load and its folder removed, or kept and used when the newer build can't be downloaded. Builds are ordered by number, not by name. The Windows self-test takes --attach for vision runs. Tested in the app with an older build installed: Mac (M1 Pro, Metal, b11853 -> b11946), Windows (RTX 4090, CUDA, b11836 -> b11946) and Linux (WSL, CUDA, b11838 -> b11946). LFM3 1.7B answered on all three (about 85, 290 and 246 tok/s); LFM3 VL 2B read a newspaper front page on Mac and Windows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Closed
5 tasks done
2 of 4 tasks
baylitoo
added a commit
to baylitoo/slm_benchmark
that referenced
this pull request
Oct 28, 2026
* feat(serving): LFM3 families (lfm3, lfm3_vl) Every LFM3 GGUF reports arch "lfm3", so the projector and Hub tags pick the family: no mmproj -> lfm3, mmproj -> lfm3_vl, an audio-only projector (LFM3-Audio-2B) is refused instead of being promoted to vision. Own KV-cache constant, 12,288 bytes/token (6 full-attention layers x 4 KV heads x head_dim 128), runtime note for ggml-org/llama.cpp#31252, and lfm3_1_7b / lfm3_vl_2b benchmark profiles. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs: LFM3 families in models/onboarding and models/reasoning Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
cubicalbatch
added a commit
to cubicalbatch/vesta
that referenced
this pull request
Oct 28, 2026
b11938 is the first release with the `lfm3` architecture (ggml-org/llama.cpp#31252, LFM3-1.7B / LFM3-VL-2B). b10373 aborts with "unknown model architecture: 'lfm3'" on those GGUFs. Verified: the llamacpp stage builds from the pinned asset; the binary loads and answers with LFM3-1.7B-Q4_K_M.gguf inside the runtime base image, and the app's router path (--models-dir + data/config/models.ini) loads LFM3 and Qwen3.5-4B unchanged.
2 of 6 tasks
3 tasks
huaxel
pushed a commit
to huaxel/strix-llama.cpp
that referenced
this pull request
Nov 1, 2026
* model : add LiquidAI/LFM3-1.7B and LiquidAI/LFM3-VL-2B Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Xuan-Son Nguyen <son@huggingface.co> Assisted-by: Claude Opus 5.5 * tests : set the SWA rope base of the LFM3 test model Assisted-by: Claude Opus 5.5 * llama : enable recurrent state rollback for LFM3 Assisted-by: Claude Opus 5.5 * Apply batched suggestions from code review Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Tarek Dakhran <github@dakhran.com> * gguf-py : align the lfm3vl mapping comments Assisted-by: Claude Opus 5.5 * llama : fix CRLF line endings from the review suggestions Assisted-by: Claude Opus 5.5 * tests : add the LFM3 Metal fusion baseline Assisted-by: Claude Opus 5.5 * llama : keep the LFM3-VL depthwise conv kernels in F32 when quantizing Assisted-by: Claude Opus 5.5 --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> Co-authored-by: Xuan-Son Nguyen <son@huggingface.co>
2 tasks done
4 of 12 tasks
Closed
3 tasks done
6 tasks
edwardyoon
added a commit
to edwardyoon/focus-llama
that referenced
this pull request
Nov 6, 2026
The fork synced ggml-org#31431 (tensor split for LFM3, which references LLM_ARCH_LFM3 in llama-model.cpp) but not ggml-org#31252 (which adds the enum value), leaving master in a non-building state. Add the enum value + name mapping (matching ggml-org#31252) to unblock the build. The full LFM3 model implementation is not included. Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Overview
Add support for the text and vision models from the LiquidAI LFM3 model family
Huge thanks to @CISC and @ngxson for contributions and review.
Additional information
Video input support will be added later.
Correctness and conformance verified against the reference HF implementation (text/chat logprobs, image sizing, tool calls, structured output, image input).
F32/Q8_0/Q4_0 quantization smoke-tested on Mac and CUDA.
Requirements