Changelog
Releases, new adapters, and maintenance windows.
Domain migration — modelforgelab.dev → modelforgelab.cloud
We have moved the primary domain from modelforgelab.dev to
modelforgelab.cloud. All existing links, tokens, and CLI configurations
pointing to the old domain continue to work via redirect for 90 days. Please update
your mfl login URL and any MFL_BASE_URL env vars at your
earliest convenience. Historical download counts and adapter metadata are preserved.
GGUF catalog expansion
Added codellama-13b-python-gguf and qwen25-32b-instruct-gguf
to the public catalog. Improved format detection in mfl pull for GGUF
artifacts. SHA-256 verification now runs in a background thread to avoid blocking
large downloads.
Community highlight: @concept_pixels's FLUX Game Concept Art crossed 34 000 downloads this month — the fastest-growing community adapter since launch.
Community spotlight — Q1 2026
Q1 brought 38 new community adapters, with highlights from @photorealist_x (FLUX Hyperrealism, 31k+ downloads), @career_forge_ai (Qwen2.5 CV & Resume Writer), @hire_coach_ai (Mistral Interview Coach), and @animeflux_dev (FLUX Anime Character XL). Combined downloads for Q1 community submissions: over 85 000.
GGUF format support
The registry now natively indexes and serves GGUF artifacts. mfl pull
detects the format and emits the correct Modelfile for Ollama. Range
requests and sha256 verification work identically to safetensors.
Maintenance window
Storage backend migrated to a new volume. Downtime 03:10–03:48 UTC. All artifact checksums re-verified after migration.
API tokens and per-adapter visibility
Members can now generate scoped API tokens (read-only or read-write) under account settings. Adapter visibility can be toggled between public, members, and private without re-uploading.
One-command installer
Single-command installer for fresh VPS deployments. Ships with a default Nginx config, systemd unit, and Let's Encrypt helper. Breaking: config file format changed from TOML to env-based; see migration guide.
Community spotlight — Summer 2025
Summer saw a surge of community contributions: @grimforge's SDXL Dark Fantasy adapter became the most-downloaded image adapter in the registry (27k downloads in three months). @jsdevbot and @stellarpixel submitted highly-rated coding and sci-fi concept art adapters. @quant_lab's GGUF Q4 quantisation of Llama 3 8B preceded official GGUF support and helped shape our format-detection pipeline.
New adapters: FLUX.1 portrait and 3D render
Two new members-only FLUX.1-dev adapters: flux-portrait-lora for
consistent studio portrait style and flux-3d-render-lora for Blender
Cycles-style output. Available under FLUX.1-dev non-commercial license.
DoRA training backend
Fine-tune jobs can now use DoRA (decomposed weight adaptation) in addition to
standard LoRA. Pass --method dora to the CLI or toggle in the UI.
Typical improvement: 1–3% on instruction-following benchmarks at the same rank.
Resumable pulls
Artifact downloads now support HTTP Range requests, so interrupted
pulls resume instead of restarting. Added llama3-8b-changelog-writer
(GGUF, q4_k_m).
Playground improvements
Playground now streams token-level progress and reuses warm workers for faster first output. Image result is returned inline over SSE; no separate polling needed.
Fine-tune queue
Queue LoRA runs directly from an uploaded dataset. Streaming logs over SSE. Jobs are retained for 7 days; resulting adapters land directly in your private registry.