Swift Qwen3.8 27B on RTX 3090 (2026): Faster Answers

Bottom Line

• Swift does not raise tok/s. It is a LoRA fine-tune merged into Qwen3.8-27B. The config, the layer count and the file sizes are the same, so decode speed on a 3090 is the same.
• It saves wall-clock time per answer. It writes fewer thinking tokens. At about 40 tok/s, the GPQA-Diamond mean falls from about 6.3 to 3.7 minutes per answer. Our arithmetic.
• “58.3% fewer” is the best single number in the table. It is the GPQA-Diamond median. The GPQA-Diamond mean is 41.0%. The coding mean is 24.3%.
• Use Q4_K_M on a 24GB card. It is 16.79 GiB. Add 64K of context and it is about 20.79 GiB. Use IQ4_XS (14.61 GiB) for 128K.
• The license has a revenue cap. Commercial use is free under US$1,000,000 gross revenue. Above that you need UkisAI’s paid Swift Enterprise License.

Every file size is a byte count from the Hugging Face API, read 2026-09-20. Every accuracy and token figure is UkisAI’s own, from BF16 or INT4 runs. UkisAI states it has not re-run the benchmark suite on these GGUF files. We did not run Swift ourselves.

What Swift Is

Swift-Qwen3.8-27B comes from UkisAI. The BF16 repo ukisai/Swift-Qwen3.8-27b was created 2026-09-08. The GGUF repo ukisai/Swift-Qwen3.8-27B-GGUF was created 2026-09-11 and last changed 2026-09-16. The API reports 136,668 downloads and 331 likes for the GGUF repo.

UkisAI found reasoning-marker tokens that, in its analysis, trigger overthinking. It fine-tuned Qwen3.8-27B to penalize those tokens. It also added a transfer component from BottleCap AI’s ThinkingCap-Qwen3.6-27B.

The NOTICE file lists what changed. The weights changed through a merged LoRA adapter. generation_config.json gained min_p: 0 and repetition_penalty: 1.0. config.json, the tokenizer and the chat template are unmodified from Qwen3.8-27B.

That last point decides the speed question. Same config means same compute per token.

The 58.3% Claim, Read Closely

This is UkisAI’s BF16 table. The base is Qwen3.8-27B BF16 at thinking effort xhigh, five seeds per model.

Benchmark	Base score	Swift score	Base mean tokens	Swift mean tokens	Mean cut	Median cut
GPQA-Diamond	88.38%	88.28%	15,014	8,855	41.0%	58.3%
MMLU-Pro	85.47%	84.95%	2,980	1,603	46.2%	28.3%
C-Eval	90.00%	90.62%	1,492	804	46.1%	19.3%
IFBench	73.53%	71.80%	8,052	4,657	42.2%	50.5%
AIME 2026	98.67%	94.00%	22,014	16,143	26.7%	50.2%
HMMT (Nov 2025)	99.33%	96.00%	22,032	15,189	31.1%	45.9%
ERQA	67.45%	66.30%	4,137	2,045	50.6%	54.6%
Terminal-Bench 2.1	66.74%	65.84%	37,086	27,272	26.5%	38.7%
LiveCodeBench v6	76.76%	81.55%	11,374	8,615	24.3%	45.8%

Three things the headline hides:

1. 58.3% appears once. It is the GPQA-Diamond median. No other cell reaches it.
2. “Less than 1% loss” does not hold for math. AIME 2026 drops 4.67 points and HMMT drops 3.33 points.
3. Coding improves, but saves the fewest tokens. LiveCodeBench v6 gains 4.79 points. Its mean token cut is the smallest in the table, 24.3%.

The effort settings matter too. At medium, the mean thinking cut is 22.7%. At low, it is 25.8%. At xhigh, it is 41.0%. If you already run the base model at medium, expect about half the headline saving.

UkisAI also compares Swift at xhigh against the base at medium on GPQA-Diamond. Base medium uses 4,451 mean tokens and scores 84.14%. Swift xhigh uses 8,855 and scores 88.28%. So Swift buys back about 4 points of accuracy for about twice the tokens of medium.

The Only Independent Test

MindStudio published a side-by-side test on 2026-09-19. It ran both models on one real coding task.

Metric	Qwen3.8-27B	Swift-Qwen3.8-27B
Output tokens	11,744	60,572
Tool calls	4	21
Time to finish	47 min 7 s	35 min 52 s

Swift wrote about five times the tokens and still finished first. The base model spent its time deliberating. This is one task, not a benchmark. It does show that “fewer thinking tokens” is a benchmark average, not a rule for agent loops.

We found no independent tok/s figure for Swift on any GPU.

GGUF File Sizes, in GiB

The card lists decimal GB. Your GPU reports GiB. These are the API byte counts, converted.

File	GiB	UkisAI KLD (wikitext @512)	Fits 24 GiB?
IQ3_XXS	11.67	0.0724	Yes
IQ4_XS	14.61	0.0165	Yes, room for 128K
Q4_K_S	15.44	0.0150	Yes
Q4_K_M	16.79	0.0120	Yes, UkisAI’s 24GB pick
Q4_K_L	17.73	0.0102	Yes
Q5_K_M	18.81	0.0056	Yes, up to about 64K
Q6_K	21.31	0.0020	Tight, about 32K
Q8_0	27.12	0.0009	No
mmproj F16 (vision)	0.86	—	Adds to any row

Lower KLD means closer to BF16. UkisAI measured it. We did not.

Every tier stores the MTP head at Q8_0. The Q4_K_M tier also lifts ssm_out, attn_gate, output and token_embd to Q6_K. UkisAI says those lifts cost about 1.1 GB and cut KL divergence by about 20% against a plain llama.cpp Q4_K_M.

Time Saved per Answer on a 3090

Swift has the same architecture as the base model. So we use the base model’s measured 3090 speed. Hardware Corner measured 40.31 tok/s at 4K and 33.95 tok/s at 64K on a 4-bit file. See our Qwen 3.8 27B on RTX 3090 page for the test details.

Benchmark (mean tokens)	Base at 40 tok/s	Swift at 40 tok/s	Saved
GPQA-Diamond (15,014 → 8,855)	6.3 min	3.7 min	2.6 min
IFBench (8,052 → 4,657)	3.4 min	1.9 min	1.4 min
LiveCodeBench v6 (11,374 → 8,615)	4.7 min	3.6 min	1.1 min
AIME 2026 (22,014 → 16,143)	9.2 min	6.7 min	2.4 min

Our arithmetic: mean tokens divided by 40 tok/s. It ignores prompt processing and assumes speed stays flat. Speed does fall with context, to about 34 tok/s at 64K, so long answers take longer than shown. Treat each row as a rough lower bound.

MTP adds to this. The base model’s MTP head lifted one community 3090 test from 41.55 to 66.39 tok/s at 8K. Swift keeps the head. Nobody has published an MTP acceptance rate for Swift, so we do not quote a Swift MTP speed.

The Thing Nobody Else Says: Context Budget

Fewer thinking tokens also means a smaller KV cache per answer. Qwen3.8-27B keeps a growing cache on only 16 of 64 layers. Its config.json gives 4 KV heads and head dim 256.

2 (K and V) x 16 layers x 4 KV heads x 256 head dim x 2 bytes = 65,536 bytes per token

That is 64 KiB per token in f16. UkisAI’s card states the same figure.

On a 24GB card, the saving is small. The GPQA-Diamond mean saves 6,159 tokens. That is about 0.38 GiB of cache per answer. Terminal-Bench saves 9,814 tokens, about 0.60 GiB. Our arithmetic. Q4_K_M at 64K already fits with room, so you will not feel this.

On a 16GB card, it decides whether the answer fits at all. UkisAI tells you to set at least 65,536 tokens of context in LM Studio, koboldcpp and Jan. A 64K f16 cache is 4.00 GiB.

16 GiB card	Weights	+ 64K cache	Fits?
IQ4_XS	14.61	18.61	No
IQ3_XXS	11.67	15.67	Only just, no buffers
IQ4_XS, 16K cache	14.61	15.61	Only just

A 16K window cannot hold a base-model AIME answer of 22,014 mean tokens. Swift’s 16,143 mean still overflows it. So on 16GB, Swift helps, but it does not fix the fit. You still drop to IQ3_XXS or quantize the cache.

One more trap: Ollama sizes the context to 32,768 tokens on a 24GB card, per UkisAI’s card. That is half the 65,536 UkisAI recommends. Set OLLAMA_CONTEXT_LENGTH=65536 or num_ctx 65536. A short window on a long reasoning trace looks like an endless loop.

Fit by Context on 24GB

Weights plus f16 KV cache. Vision adds 0.86 GiB.

File	32K	64K	128K
IQ4_XS (14.61)	16.61	18.61	22.61
Q4_K_M (16.79)	18.79	20.79	24.79, no
Q5_K_M (18.81)	20.81	22.81	No
Q6_K (21.31)	23.31, tight	No	No

These sums exclude compute buffers and the linear-attention state. Treat each one as a lower bound.

What the License Allows

The weights ship under the Swift Open License v1.0. We read the full LICENSE file on 2026-09-20. In plain terms:

• Free: personal, research, educational and evaluation use, for anyone.
• Free commercial use: if you and every company you control, or that controls you, had under US$1,000,000 gross revenue in the last completed fiscal year.
• Not licensed: commercial use by any entity at or above that threshold. It needs the paid Swift Enterprise License. Contact UkisAI for terms. No price is published.
• Non-profits: a qualified 501(c)(3) or foreign equivalent is exempt from the threshold for non-commercial or research use.
• Redistribution: allowed, including quants. You must pass on the license, mark changed files, keep the NOTICE, and include the Apache 2.0 license for the base model.
• Termination: the license ends automatically if you break any term. You must then delete Swift. Your Apache 2.0 rights in base Qwen3.8-27B stay intact.

Note one mismatch. The GGUF card says “annual recurring revenue.” The LICENSE file says gross revenue. The LICENSE file is the binding text. Gross revenue is the stricter test.

If your company is above the threshold, run base Qwen3.8-27B instead. It is Apache 2.0 with no revenue cap.

Which GPU to Buy

Prices are US street ranges as of August 2026, from our hardware price reference. The 2026 DRAM shortage moved them hard. Check current listings before you buy.

The used RTX 3090 is the card for this model. It holds Q4_K_M with 64K of context, or IQ4_XS with 128K. Used listings ran $1,000–1,300 (as of August 2026). One scraped tracker showed $1,399 on 2026-09-16, so the range is drifting up.

Check RTX 3090 24GB prices on Amazon

The RTX 4090 gives the same fit, faster. It has the same 24GB, so the table above does not change. It costs about $2,270–2,600 used (as of August 2026). Buy it for long prompts, where prompt processing dominates. See RTX 3090 vs RTX 4090.

Check RTX 4090 24GB prices on Amazon

A 16GB card runs Swift only with compromises. The RTX 5060 Ti 16GB sits at $589–805 (as of August 2026). It needs IQ3_XXS for UkisAI’s 64K minimum, or a quantized cache. Swift’s shorter answers help most here, but they do not remove the squeeze.

Check RTX 5060 Ti 16GB prices on Amazon

How to Run It

UkisAI’s llama.cpp command, with context cut to fit 24GB:

llama-server -hf ukisai/Swift-Qwen3.8-27B-GGUF:Q4_K_M \
  --jinja -fa on -ngl 99 -c 65536 \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0 \
  --presence-penalty 0 --repeat-penalty 1.0 \
  --spec-type draft-mtp --spec-draft-n-max 3

-hf also pulls the vision projector. Drop the last line to turn MTP off. For 128K, swap Q4_K_M for IQ4_XS and set -c 131072.

Sources

• Hugging Face API, ukisai/Swift-Qwen3.8-27B-GGUF?blobs=true (createdAt 2026-09-11, lastModified 2026-09-16, downloads, likes, all file byte counts), read 2026-09-20
• ukisai/Swift-Qwen3.8-27B-GGUF README.md (KLD table, recipe, KV cache, Ollama and context notes, validation scope), read 2026-09-20
• ukisai/Swift-Qwen3.8-27b README.md (BF16 table, reasoning-effort table, INT4 table), NOTICE and LICENSE, read 2026-09-20
• Qwen/Qwen3.8-27B config.json (64 layers, full_attention_interval: 4, 4 KV heads, head dim 256, 262,144 max positions), read 2026-09-20
• MindStudio, “Qwen 3.8 27B Swift vs Original: Does It Really Think Less?”, 2026-09-19
• Base-model 3090 speeds as cited on our Qwen 3.8 27B on RTX 3090 page (Hardware Corner; thc1006 community test)
• OpenClaw DC hardware price reference, compiled 2026-08-11, last updated 2026-09-16

See Also

• Qwen 3.8 27B on RTX 3090 — measured base-model speed and the MTP gain on 24GB
• Qwen 3.8 27B GSQ-RCO quants — the same base model at 2.50 to 3.50 bits for 12GB and 16GB cards
• Best local LLM for 16GB VRAM — what fits when Swift does not
• How much VRAM for 128K context — the cache arithmetic across model sizes
• RTX 3090 vs RTX 4090 for local LLMs