license: other
license_name: swift-open-license-1.0
license_link: LICENSE
base_model: d0xin/Swift-1.5-Qwen3.8-27B-Uncensored-BF16
base_model_relation: quantized
library_name: llama.cpp
pipeline_tag: text-generation
tags:
- gguf
- quantized
- qwen3.8
- swift-1.5
- uncensored
- mtp
- 8gb
- llamampere
- ampere
Swift 1.5 Qwen3.8-27B Uncensored, ATX AI2 8 GB (2.09 bits per weight on card, with MTP)
8 GB cards need llamAmpere v0.7.1 (
--preset 8dense; coming, not yet released). The file already runs on llamAmpere v0.7.0.1 with--preset 10dense(10 GB cards). Stock llama.cpp cannot load it.
Our 8 GB build of Swift 1.5 27B: the 27B model's weights in 6.2 GiB of card memory (the 1.3 GiB token-embedding table stays in system RAM), plus a 0.12 GiB MTP draft layer, with 204,800 tokens of context on an 8 GB RTX 30-series card. MTP (multi-token prediction) lets the model draft several tokens at once and check them in one pass, which speeds up writing.
| File | Size | SHA-256 |
|---|---|---|
qwen3.8-27b-Swift1.5-Uncensored-ATX-AI2-8GB.gguf |
8.39 GB (8,393,828,704 bytes) | 384237ca464ffe3c80e24119479afd3fe0e5934667068de601f18fb4d2b82f56 |
lsa-qwen38-27b-d2048.gguf (sparse-attention map, needed by the presets) |
8,390,624 bytes | 5489a4f2f6e9169c57db7344afdebdca243c3d438d8e70197c52a6c95244aa68 |
mtp-dm.gguf (small MTP draft layer for 8 GB cards) |
127,489,024 bytes | 4a8e9a5cfb6b1395088bd88b26400b504e47a10fad5a109e4cbee3276f9d5cae |
Keep all three files in one folder.
Quality
Drift from the full-precision model (KLD, lower is better). At each token position we compare the quant's full next-token distribution with Swift 1.5 BF16's. "Top-1" is how often both pick the same next token. "Bits per weight" counts the main model and output head that sit on the card; the token embedding table lives in system RAM and is not counted.
Test set 1: fresh text (18,000 positions). Six coding, agent-session and documentation prompts built from public material published after Swift 1.5 and after our calibration data (3,000 scored tokens each, read at 6,000 to 8,700 tokens of context). None of this text was seen during quantization or healing.
| File | Card memory (weights) | Bits per weight | KLD | Coding | Agentic | Docs | Top-1 |
|---|---|---|---|---|---|---|---|
| This model | 6.22 GiB | 2.09 | 0.0857 | 0.0999 | 0.0871 | 0.0701 | 92.85% |
| Same file before healing (see below) | 6.22 GiB | 2.09 | 0.0975 | 0.1135 | 0.0994 | 0.0795 | 92.35% |
| Unsloth Qwen3.8-27B UD-IQ2_XS | 6.79 GiB | 2.28 | 0.0924 | 0.1100 | 0.0925 | 0.0747 | 92.52% |
| Unsloth Qwen3.8-27B UD-IQ2_XXS | 6.40 GiB | 2.14 | 0.1130 | 0.1335 | 0.1123 | 0.0934 | 91.80% |
Test set 2 (16,200 positions: 5,400 each of coding, agentic and documentation text, read at about 12,000 tokens of context):
| File | Card memory (weights) | Bits per weight | KLD | Coding | Agentic | Docs | Top-1 |
|---|---|---|---|---|---|---|---|
| This model | 6.22 GiB | 2.09 | 0.0868 | 0.0899 | 0.0999 | 0.0707 | 92.52% |
| Same file before healing | 6.22 GiB | 2.09 | 0.0994 | 0.1020 | 0.1130 | 0.0834 | 91.93% |
| Unsloth Qwen3.8-27B UD-IQ2_XS | 6.79 GiB | 2.28 | 0.0935 | 0.0900 | 0.1062 | 0.0843 | 92.02% |
| Unsloth Qwen3.8-27B UD-IQ2_XXS | 6.40 GiB | 2.14 | 0.1135 | 0.1081 | 0.1307 | 0.1019 | 91.44% |
| Our AI2 2.3 bpw (10 GB cards) | 6.92 GiB | 2.32 | 0.0593 | 0.0631 | 0.0684 | 0.0465 | 94.06% |
AI2 numbers are re-measured with this card's method (weights on the card only), so they differ slightly from the AI2 card.
What the Unsloth rows mean. The Unsloth files are quantizations of stock Qwen3.8-27B, and here they are scored against Swift 1.5 BF16, a fine-tuned model. So these tables measure how close each file is to Swift 1.5 BF16, not how good each quantizer is: part of the Unsloth files' KLD is the difference between stock Qwen and Swift 1.5. Scored against its own parent (stock Qwen) on the test set 2 text, UD-IQ2_XS has a KLD of 0.074, lower than this file's KLD against Swift 1.5. What the tables do show: for someone who wants Swift 1.5, this file stays closer to Swift 1.5 BF16 than either stock file (about 7% lower KLD than UD-IQ2_XS on both test sets) while using 8% less card memory than UD-IQ2_XS and 3% less than UD-IQ2_XXS.
Against quants we built ourselves from the same Swift 1.5 BF16 with stock llama-quantize and its importance matrix (test set 2, fp16 KV cache, from our AI2 model card): IQ2_XS (2.78 bpw) 0.1430, IQ2_S (2.82 bpw) 0.1245, Q2_K (3.28 bpw) 0.0801.
LiveCodeBench v6 (100 problems, 3 runs each, temperature 1.0, reasoning effort xhigh, no output cap other than the context window):
| File | Problems solved (of 300) |
|---|---|
| This model | 244 (81.3%) |
| Same file before healing | 239 (79.7%) |
Healing did not hurt coding: +1.7 problems per run on average (paired over the same problems and runs, 95% range −2.0 to +5.3), which is within noise. Pass/fail counts each answer's code against the problem's own tests.
How it was made
- Natively encoded from the Swift 1.5 BF16 checkpoint in our AI+ format. Most of the weights (16.4 billion) use a 1.78-bit type; the rest use 2.0 to 5.1 bits, chosen by how much each tensor matters. The output head is 2.5-bit.
- After encoding we ran a short healing step. All quantized codes stayed frozen. Only the fp16 scales already in the file and the RMSNorm weights were adjusted (58.7 million numbers out of 27 billion), by matching the BF16 model's output probabilities on a small calibration set. On the fresh text, healing lowered KLD from 0.0975 to 0.0857 (−12.1%, 95% range −13.4% to −10.8%), lowered it in all 6 sequences, and raised top-1 from 92.35% to 92.85%. On test set 2 it was −12.7% (0.0994 to 0.0868).
- This is not coding post-training. The calibration recipe is the same as for our AI2 file's healing step: a small mixed set of prompts answered by the BF16 model. It shares no documents with either test set (one short markdown table line is the only repeated 13-token sequence), and no LiveCodeBench problems were used. The gain on never-seen fresh text matches the gain on the calibration set's own held-out part (−13.8%), so healing is not fitted to the test text.
- File size and tensor layout are unchanged by healing, so speed should be too (the speeds below were measured before healing).
The GGUF contains the text model and its own MTP draft layer. It does not contain the vision tower. This is an uncensored, refusal-reduced derivative of UkisAI's Swift 1.5, built on Qwen3.8-27B.
Speed and memory
Measured on a rented RTX 3070 8 GB (PCIe 4.0 x16), llamAmpere v0.7.1 --preset 8dense (context 204,800, mtp-dm.gguf draft layer, MTP on), chat prompt, temperature 1.0, 5,120-token replies. These runs used the file before healing, which has the same size and layout.
| Prompt already in context | Writing speed |
|---|---|
| 100,000 tokens (6 runs) | 41.85 tokens/s |
| 199,400 tokens (3 runs) | 36.53 tokens/s |
Peak card memory was 7,607 MiB across these runs.
Card memory stays under 7,692 MiB (8 GB minus 500 MiB), so 204,800 tokens fit on a card that is not also driving a display. On a 10 GB card, --preset 10dense peaked at 8,945 MiB with its default 180,224-token context (RTX 3090, llamAmpere v0.7).
Run it
Get llamAmpere v0.7.1 (coming, not yet released; v0.7.0.1 works for the 10 GB command below). Linux or Windows, NVIDIA RTX 30-series. Then:
hf download jakeatx/qwen3.8-27b-Swift1.5-Uncensored-ATX-AI2-8GB --local-dir ~/models/atx-ai2-8gb
# 8 GB card, 204,800 tokens of context
llama-server -m ~/models/atx-ai2-8gb/qwen3.8-27b-Swift1.5-Uncensored-ATX-AI2-8GB.gguf \
--preset 8dense --host 127.0.0.1 --port 8080 --jinja --reasoning-format deepseek
# 10 GB card (also works on v0.7.0.1)
llama-server -m ~/models/atx-ai2-8gb/qwen3.8-27b-Swift1.5-Uncensored-ATX-AI2-8GB.gguf \
--preset 10dense --host 127.0.0.1 --port 8080 --jinja --reasoning-format deepseek
Recommended sampling: temperature 1.0, top-k 20, top-p 0.95, min-p 0. Reasoning effort is chosen per request with chat_template_kwargs: {"reasoning_effort": "low" | "medium" | "xhigh"} (default xhigh), or {"enable_thinking": false} to turn thinking off.
Coding agents (pi, OpenCode, Hermes Agent): ready-made settings for this server are in jakeatx/qwen3.8-27b-Swift1.5-ATX-8GB-agent-adapter.
License and credits
Swift Open License 1.0, inherited from Swift 1.5; see LICENSE, LICENSE-APACHE-2.0 and NOTICE. Base model: Qwen3.8-27B by Alibaba Cloud (Apache 2.0), fine-tuned as Swift 1.5 by UkisAI, uncensored BF16 by d0xin. Runtime: llamAmpere, a fork of llama.cpp (MIT) by Georgi Gerganov and contributors.