Almost Opus 4.6 Max quality with JiRack DeltaNet 27B — but it runs on your PC
CMSManhattan/JiRackDeltaNet_27b is a 27B hybrid Qwen3.8 / Qwen3.5-style DeltaNet model on Hugging Face. It is positioned for CPU inference, tool calling, routing, and robotics-oriented tags.
Repo size is listed at about 256 GB. The README was last updated about 19 hours ago. Model weights are MIT.
What it is
The model card describes it as Qwen 3.8 27B migrated onto a DeltaNet architecture: hybrid full attention plus Gated DeltaNet / SSM-style layers.
It uses an extended JiRackDeltaNetTokenizer with special tags for routing, media, vision, sound, tool calls, and robotics. Context is advertised up to 262,144 tokens.
Architecture notes from the model card
- Hidden size: 5120
- Layers: 65
- Attention: 24 query heads / 4 KV heads (GQA)
- FFN intermediate: 17408
- SSM: conv kernel 4, state 128, group count 16, time-step rank 48, inner size 6144
- RoPE θ: 10,000,000
That matches the public Qwen3.x-27B hybrid DeltaNet layout used elsewhere: 48 Gated-DeltaNet layers, periodic full-attention layers, and MTP-style extra attention in some engines.
Repo tags mix deltanet, qwen3.8, and qwen3_5_text with ternary, bitnet, and jirack. The README still frames it as a CPU / QAT-friendly JiRack model.
The published GGUF line is standard llama.cpp-style quants from FP16 (Q2_K, Q3_K_M, Q4_K_M, plus Q6 / Q8 on Ollama). That is not a separate 1.58-bit BitNet checkpoint in the card text.
How to run it
Ollama
Reasoning is off by default in the recommended tag:
ollama pull cmsmanhattan/JiRackDeltaNet_27b-q4-reasoning
ollama run cmsmanhattan/JiRackDeltaNet_27b-q4-reasoning "What is the capital of France?"
Override with --think=false.
Org page: ollama.com/cmsmanhattan. Tags include q2 / q3 / q4 / q6 / q8 and full reasoning variants, roughly 10–54 GB. Pull counts there are still very low (1–2).
Docker
UI on http://localhost:7869:
docker run -d --name jirack_deltanet_27b -p 7869:7869 --restart unless-stopped \
cmsmanhattan/jirack_deltanet_27b-cpu-q4:latest
The card lists Q2 / Q3 / Q4 / Q6 / Q8 / full images. 27B Docker builds are also noted as available on request.
Quants and RAM
| Variant | Approx. size | Suggested RAM |
|---|---|---|
| Full | ~55 GB | 48–64 GB |
| Q8 | ~29 GB | higher-end workstation |
| Q6 | ~22 GB | 24 GB+ comfortable |
| Q4_K_M | ~16.8 GB | 18–24 GB (recommended) |
| Q3_K_M | ~13.9 GB | 15–20 GB |
| Q2_K | ~11.2 GB | 12–17 GB |
Recommended box: Ryzen 9 / i9 / Xeon, 24–32 GB RAM, Q4_K_M. Extra RAM is for KV cache and long context, not just weights.
Licensing
Read the card carefully.
- Model weights: MIT. Use, modify, and commercially redistribute the weights.
- Docker image, UI, and prebuilt Ollama packaging are called out separately as commercial ($12 per user per year in the card).
- UI clients are free when used with official JiRack Docker containers, not if redistributed alone.
Related pieces
- Tokenizer: CMSManhattan/JiRackDeltaNetTokenizer
- Tool-call notes point at Spring AI Alibaba, Gorilla / GoEx, ToolBench / Hermes-style function-calling data, plus JiRack tags
- Benchmarks in the card are upstream Qwen3.8-27B numbers (SWE-bench Pro, Terminal Bench, GPQA, OSWorld, and others), not JiRack-specific post-train or quant scores
This 27B line sits on the same JiRack research track as the ternary 1B work: CMSManhattan/JiRackTernaryPro_1b and kgrabko/JiRackTernary_1b.
If you want the next step — GGUF file list and exact shard names, a local llama.cpp command, an Ollama Modelfile, or a comparison versus JiRack Ultra / ternary 1B–32B — say which runtime you are targeting.
