What GPU do you need to run Swift 1.5 Qwen3.8-Flash-Next?
A new Framework Desktop ($6,799). No priced GPU card holds Swift 1.5 Qwen3.8-Flash-Next (about 121.1 GB at Q4); this system holds it with room for its full 256K context, at about 42 tok/s estimate. 2 of the 157 GPUs and systems we track fit it.
180B parameters, 6B active (mixture-of-experts) (Hugging Face, accessed 2026-11-11). Q4_K_M GGUF: 119.6 GB (source). Architecture config (Hugging Face config.json): 12 attention layers (the rest are recurrent) × 2 KV heads × 256 head dim, so about 24 KB of FP16 KV cache per token of context (how we estimate context).
UkisAI's reasoning-efficient fine-tune of Qwen3.8-Flash-Next: shorter thinking traces, tuned for coding and long agent tasks. Its config.json is identical to the base model's, so 6B active is Qwen's published count for the base; UkisAI's card does not restate it. UkisAI claims "63.4% fewer thinking tokens, with a 1.8x speed up while keeping the accuracy loss <1% vs base on xhigh"; the 1.8x is fewer tokens to finish one demo prompt, not faster decoding (details below). The checkpoint keeps the base model's 1-layer MTP head, so MTP self-speculative decoding is possible; that is not a speed figure. Q4_K_M is UkisAI's official GGUF (3 parts, 119,599,029,952 bytes); the vision projector is extra. License: free for commercial use only up to US$1,000,000 gross annual revenue (with affiliates), and the Qwen Community License 1.0 also applies (details below). Release date is the Hugging Face repo creation date.
What we checked
Checked Nov 11, 2026. Official is the maker or the weight files. A vendor claim is that project's own documentation. A community claim is someone else's upload. Our check is a file or search we read. Unknown means we have no source, so it is not estimated. None of these rows is a speed we measured.
| Kind | What | Detail | Source |
|---|---|---|---|
| Official | License | UkisAI's contribution is under the Swift Open License v1.0. The card says it is free for personal, research, educational, evaluation and commercial use for individuals and organizations with gross annual revenue (including affiliates) up to US$1,000,000; above that, commercial use needs a separate Swift Enterprise License. This is a summary of the license, not legal advice. | LICENSE, checked Nov 11, 2026 |
| Official | Base model license | The base model's Qwen Community License 1.0 still applies. Under it, organizations running a Model-as-a-Service or AI Work Assistant business need a separate license from Qwen before commercial use, and products above 100 million monthly active users or US$20 million monthly revenue must display the model name. This is a summary of the license, not legal advice. | LICENSE-QWEN, checked Nov 11, 2026 |
| Official | Architecture | config.json is identical to Qwen/Qwen3.8-Flash-Next's: model_type qwen4_exp, 48 layers (12 full attention), 2 KV heads, head size 256, 512 experts with 10 routed per token, 1 MTP layer and a vision tower. Hugging Face's safetensors total is 179,999,981,424 parameters, the same as the base. 6B active is Qwen's published figure for the base model. The card mentions YaRN scaling to 1,000,000 tokens, following Qwen's card; this page uses the config's 262,144. | config.json, checked Nov 11, 2026 |
| Official | Weight files | 131 BF16 safetensors shards sum to 360,000,192,888 bytes (360.00 GB) at commit 0bd4fe22431372cdad1979267d3ab45aa7e6150a. UkisAI: "The BF16 checkpoint is about 360 GB, so it needs a multi-GPU node. We evaluated it with tensor parallelism 8." These are downloads, not the memory a runtime needs. The repo was renamed from ukisai/Swift1.5-Qwen3.8-Flash-Next, which redirects. | Hugging Face API, checked Nov 11, 2026 |
| Official | GGUF files | UkisAI's standard GGUF repo (commit 0142554c963a8653912c0c8283f05f13f63653f3, architecture qwen4exp) has the Q4_K_M in 3 parts: 44,648,342,368 + 44,928,589,600 + 30,022,097,984 bytes. A second repo, Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF (commit 99bb8f7f95c7aa7b24a36a7786a4f657b30f5d3d, created 2026-09-24, last modified 2026-10-07), has compact GSQ-RCO files. UkisAI says the GSQ-RCO tier names describe mixed-precision allocation profiles, not plain llama.cpp quants. It labels Q2_0 experimental and says "For a file of similar size, prefer IQ2_XS". Every size is in the table below. | GGUF repo, checked Nov 11, 2026 |
| Vendor claim | UkisAI's efficiency claim | UkisAI's card: "63.4% fewer thinking tokens, with a 1.8x speed up while keeping the accuracy loss <1% vs base on xhigh". The 63.4% is the median thinking-token cut on GPQA-Diamond at reasoning effort xhigh (the mean cut there is 55.8%); at medium and low effort the score drops 2.62 and 2.42 points. The 1.8x comes from one demo prompt: 8 min 52 s for the base and 4 min 56 s for Swift to build a game. That is fewer tokens to finish, not faster decoding. UkisAI measured its scores in BF16 with MTP disabled. These are UkisAI's claims, not checked by us. | Model card, checked Nov 11, 2026 |
| Vendor claim | MTP drafting | The checkpoint keeps the base model's 1-layer MTP head, and the card gives vLLM and SGLang flags for MTP self-speculative decoding. That makes drafting possible; it is not a speed figure, and the speed estimates on this page do not assume it. | Model card, checked Nov 11, 2026 |
| Community claim | bartowski GGUF | bartowski/ukisai_Swift-Qwen3.8-Flash-Next-GGUF also exists. We did not size it; this page uses UkisAI's own files. | bartowski repo, checked Nov 11, 2026 |
| Our check | MLX (Mac) | There is no official UkisAI MLX build of Swift Flash Next. UkisAI's Swift-1.5-4bit-MLX is the 27B model, not this one. The community builds in the next two rows are unchecked by us. | UkisAI on Hugging Face, checked Nov 11, 2026 |
| Community claim | Dankpaws MLX build | Dankpaws/Swift1.5-Qwen3.8-Flash-Next-MLX-4.7bpw is 106,723,353,968 bytes in total. Its card (the uploader's claim) says it targets a 96 GB Mac Studio with M5 Ultra, runs on mlx-serve 26.10.1 rather than stock mlx-lm, is text only, and keeps 32.0 GB of that as an n-gram table memory-mapped on disk (74.7 GB of weight shards). We have not run it and do not list its speed table; Mac fits on this page come only from our estimate for the Q4_K_M file. | Dankpaws card, checked Nov 11, 2026 |
| Community claim | oMLX oQ builds | arkham00 publishes oMLX "oQ" mixed-precision builds, such as Swift-Qwen3.8-Flash-Next-oQ3.5e-mtp (89,777,274,008 bytes) and Swift-Qwen3.8-Flash-Next-oQ4e-mtp (106,275,556,357 bytes). Those are the uploader's builds. We have not run them and do not list their speed tables. | arkham00 card, checked Nov 11, 2026 |
GGUF files and sizes
Exact sizes from Hugging Face, accessed 2026-11-11. The fit and context tables on this page use the Q4_K_M file (119.6 GB); a smaller file needs less memory by the difference. VRAM needed is the file + the same headroom.
| File | Published by | Bytes | Size | VRAM needed | Source |
|---|---|---|---|---|---|
| GSQ-RCO Q2_0 (mixed precision, experimental) | UkisAI | 66,549,952,800 | 66.55 GB | ~68 GB | Swift-Qwen3.8-Flash-Next-GSQ-RCO-Q2_0 (2 parts) |
| GSQ-RCO IQ2_XS (mixed precision) | UkisAI | 68,152,167,168 | 68.15 GB | ~69.7 GB | Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ2_XS (2 parts) |
| GSQ-RCO IQ3_XXS (mixed precision) | UkisAI | 75,966,073,120 | 75.97 GB | ~77.5 GB | Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS (2 parts) |
| GSQ-RCO IQ3_S (mixed precision) | UkisAI | 83,743,737,120 | 83.74 GB | ~85.2 GB | Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S (2 parts) |
| Q4_K_M | UkisAI | 119,599,029,952 | 119.6 GB | ~121.1 GB | Q4_K_M/Swift-1.5-Qwen3.8-Flash-Next-Q4_K_M (3 parts) |
| mmproj-BF16 Vision projector, extra | UkisAI | 907,543,520 | 0.91 GB | — | mmproj-Swift-Qwen3.8-Flash-Next-BF16.gguf |
Other places to get it
Checked Nov 14, 2026. We list only copies we opened and checked. Check file hashes for copies from other uploaders.
- ModelScope: Other uploaderSame account name as on Hugging Face; we found no link from the lab to it
Single GPUs that fit, cheapest first
| Single GPU that fits | VRAM | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | 192 GB | ~42 tok/s | 256K (full) | low $6,799 |
| NVIDIA DGX Spark (128GB) | 128 GB | ~48 tok/s | 256K (full) | low $6,949.99 |
Every device that runs it
Graphics cards, Macs and mini PCs from the wider hardware list behind My rig, smallest usable memory first. Not sure about yours? Check your exact rig.
Show all 10 devices that run it
| Device | Type | Memory | Speed estimate | Max context estimate | Price |
|---|---|---|---|---|---|
| AMD Ryzen AI Max 385 (128GB) | Mini PC | 128 GB | ~40 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max 390 (128GB) | Mini PC | 128 GB | ~40 tok/s | 256K (full) | No price yet |
| AMD Ryzen AI Max+ 395 (128GB) | Mini PC | 128 GB | ~40 tok/s | 256K (full) | low $3,499.99 |
| NVIDIA DGX Spark (128GB) | Mini PC | 128 GB | ~48 tok/s | 256K (full) | low $6,949.99 |
| NVIDIA RTX Spark N1X laptop (6144-core GPU, 128GB) | Laptop | 128 GB | ~52 tok/s | 256K (full) | No price yet |
| Apple M2 Ultra 192GB | Mac | 192 GB (~144 usable) | ~76 tok/s | 256K (full) | No price yet |
| Apple M3 Ultra 256GB | Mac | 256 GB (~192 usable) | ~77 tok/s | 256K (full) | No price yet |
| Apple M5 Ultra 256GB | Mac | 256 GB (~192 usable) | ~89 tok/s | 256K (full) | No price yet |
| Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | Mini PC | 192 GB | ~42 tok/s | 256K (full) | low $6,799 |
| Apple M5 Ultra 512GB | Mac | 512 GB (~384 usable) | ~89 tok/s | 256K (full) | No price yet |
Multi-GPU: cards that need two or more
| Cards | Total VRAM | Speed estimate | Max context estimate |
|---|---|---|---|
| 4 × NVIDIA GeForce RTX 5090 | 128 GB | ~151 tok/s | 256K (full) |
| 3 × NVIDIA RTX A6000 | 144 GB | ~100 tok/s | 256K (full) |
Speed calculator
estimateEstimate = bandwidth ÷ GB read per token (the model's listed GGUF file: Q4_K_M for most models, otherwise the quant named in the list; scaled by active/total parameters for MoE), scaled by a per-platform efficiency and fixed per-token costs calibrated to 45 measured llama.cpp benchmarks (2025–26). Fit = file size + 1.5 GB headroom ≤ total VRAM (on Macs, the part of unified memory macOS lets the GPU use by default), and room for at least 4K tokens of context with an 8-bit KV cache. For a model with no KV cache data, only the file and headroom are checked, so a fit says nothing about context. Multi-GPU: VRAM adds, speed stays one card's (a layer split runs the cards in turn); max context uses the total. Speed is for short context. Max context = what is left after the file and a 1 GB runtime reserve ÷ KV cache bytes per token, default FP16 and 8-bit (q8_0) cache. How speed works · how context works