Qwen 3.8-Flash-Next on a 128 GB desktop: does it run, how fast, and what breaks at 1-bit

Three questions: does it run on 128 GB, how fast, and what breaks as the file shrinks. The 27B is the reference, not a rival. Six builds, four one-shot tests, thinking medium, same samplers and seed. Every leg finished every test, no crashes, no cut-offs.

• It all fits. Peak system memory: 1-bit 82 GB, 2-bit 86, 3-bit 97, IQ4_XS 99, with a 131k context. The next file (111 GB) does not fit. IQ4_XS is the top of the ladder on this box.
• Speed barely moves. 23 tok/s at 1-bit and 2-bit, 20 at 3-bit and IQ4_XS. 6B active parameters means the per-token work is the same at every quant. The 27B 4-bit does about 20 on its own backend; BF16 does 8 to 9.
• Smaller file, more thinking. Total thinking time: 1-bit 14m, 2-bit 8m 40s, 3-bit 11m, IQ4_XS 4m 23s. The 1-bit spent 75% of Blind Artist thinking, IQ4_XS 19%. That is why the 1-bit exam (25m) was slower than IQ4_XS (20m) on a smaller file.
• No visible quality loss down to 1-bit. All six Ledgers exactly right. All six drawings complete. All six games play. The only bug in 24 outputs is the IQ4_XS Ledger chart.
• Nobody signed "Qwen". Asked to sign with its own name, every Flash-Next build wrote Claude (the 1-bit: "claude-sonnet-4-6"), the 27B BF16 wrote Claude, the 27B 4-bit wrote Gemini. Same at every quant, so it is training data, not compression.

The problem

Flash-Next is a new architecture. The night of this run llama.cpp did not know it yet: Unsloth's prebuilt (b10472) refused the file, and the support lived in an open PR from Unsloth's Daniel Han, llama.cpp PR #27742, "model: add Qwen3.8-Flash-Next (qwen4exp)". So I built llama.cpp from the PR branch.

Update, Aug 27: the PR merged into llama.cpp master the next day, so a current mainline build loads the model with no tricks. If you are reading this later, skip step 1's fetch/checkout and build master; the rest of the recipe is unchanged. MTP was still marked WIP at merge time.

The recipe

On the GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB, Radeon 8060S), CachyOS. Commit 213df585b, build b10667.

1. Get the PR's code.

git clone https://github.com/ggml-org/llama.cpp tools/llama.cpp-qwen4exp
cd tools/llama.cpp-qwen4exp
git fetch origin pull/27742/head:pr-27742
git checkout pr-27742

2. Build tools. The first build failed on the missing spirv-headers; install, rebuild.

sudo pacman -S cmake vulkan-headers spirv-headers vulkan-radeon
vulkaninfo --summary   # AMD Radeon 8060S Graphics (RADV STRIX_HALO)

3. Build with Vulkan. Only the binaries the run needs. Build log linked below.

cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j $(nproc) --target llama-server llama-cli llama-gguf-split

4. Download the quants from unsloth/Qwen3.8-Flash-Next-GGUF, three shards each, about 42 minutes per file at 25 MB/s.

hf download unsloth/Qwen3.8-Flash-Next-GGUF --include "UD-IQ1_M/*"
hf download unsloth/Qwen3.8-Flash-Next-GGUF --include "UD-Q2_K_XL/*"
hf download unsloth/Qwen3.8-Flash-Next-GGUF --include "UD-Q3_K_XL/*"
hf download unsloth/Qwen3.8-Flash-Next-GGUF --include "UD-IQ4_XS/*"

5. Smoke test. Point the server at shard 1 (it finds the rest). Loaded in 19 s, answered "OK" with a reasoning field, 23.9 tok/s.

build/bin/llama-server -m <path>/UD-IQ1_M/Qwen3.8-Flash-Next-UD-IQ1_M-00001-of-00003.gguf \
  --host 127.0.0.1 --port 8898 -np 1 -c 8192 -ngl 999 --jinja --no-webui
curl -s http://127.0.0.1:8898/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"messages":[{"role":"user","content":"Say OK."}],"max_tokens":16,"temperature":0}'

6. The real launch. One server per build, full context, thinking on at medium, flash attention on. Prompts sent one shot to /v1/chat/completions with the card's samplers (1.0 / 0.95 / 20 / 0), seed 42, 120k output budget.

build/bin/llama-server -m <first shard>.gguf --host 127.0.0.1 --port 8899 -np 1 -c 131072 -ngl 999 \
  --jinja --no-webui --flash-attn on \
  --chat-template-kwargs '{"enable_thinking": true, "reasoning_effort": "medium"}'

Gotchas: this build says --chat-template-kwargs enable_thinking is deprecated (use --reasoning on on newer builds; it still worked). MTP is off, so every Flash-Next speed here is plain decoding; MTP can only make it faster.

7. Unattended. A watcher waited for the build and downloads, ran the smoke test, launched the runner with nohup. Scripts and logs linked below. Four Flash-Next legs: 1.6 hours. The 27B BF16 leg: 48 minutes.

ROCm vs Vulkan, like I'm five

Your GPU is a factory of tiny workers. You need a language to hand them work. On AMD there are two.

ROCm is AMD's own language, the cousin of Nvidia's CUDA. Fastest when it works, because AMD tuned it for their chips. The catch: you need AMD's toolkit installed, your chip has to be on AMD's list, and a source build needs the right chip flag. Unsloth's prebuilt uses it; that is the "ROCm" in the 27B rows.

Vulkan is the public road. Every GPU driver already speaks it because games use it. On Linux the AMD driver is RADV (Mesa), open source, already installed. No toolkit, one cmake flag, works on AMD, Nvidia and Intel alike. Usually a bit slower on heavy matrix math.

For a source build of an unmerged PR at 10 pm, Vulkan was the obvious pick. It cost little here: with 6B active parameters the run is limited by memory reads, not matrix math, so a 125B file on the "slow" backend ran at 20 to 23 tok/s while the 27B on the "fast" one sat at 20. Rule on this page: speeds are named with their backend, never turned into a ratio.

ROCm build, if you want it: cmake -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1151 -DCMAKE_BUILD_TYPE=Release (gfx1151 = Strix Halo). Not run here.

Create a complete falling-blocks puzzle game in a single self-contained HTML file. No external libraries, no internet resources, no images: everything inline (HTML, CSS, JavaScript, canvas rendering).

Requirements:

- 10 wide by 20 tall playfield rendered on a canvas.
- Seven distinct piece shapes, each made of four squares, each with its own color.
- Controls: left/right arrows move, up arrow rotates, down arrow soft-drops, spacebar hard-drops.
- Rotation must respect walls and stacked blocks (no clipping through anything).
- Completed horizontal lines clear, rows above fall down, and clearing multiple lines at once scores more.
- Score, lines cleared, and level displayed. Speed increases with level.
- Next-piece preview box.
- Game over when the stack reaches the top, with a visible game-over state and a restart key.
- Clean, readable dark visual style.

Output only the complete HTML file.

Create a 2D physics simulation of a volcano erupting, in a single self-contained HTML file. No external libraries: everything inline, rendered on a canvas.

Requirements:

- Side view: a volcano cone with a crater, ground, and sky.
- A pressure meter that builds over time; when it maxes out, the volcano erupts, then pressure resets and builds again on its own.
- Eruptions launch glowing lava particles from the crater with varied velocities. Particles follow gravity in a realistic arc, land on the slopes, and slide or settle downhill.
- Lava cools over its lifetime: bright yellow-orange when launched, fading to dark rock, then disappearing.
- An ash cloud of smoke particles that rises from the crater and drifts with the wind.
- Three sliders the user can move while it runs: eruption power, wind strength and direction, and particle amount.
- A live on-screen counter showing active particle count and frames per second. The simulation must stay smooth with at least 1000 active particles.

Output only the complete HTML file.

Below is a messy sales CSV export. Build a single self-contained HTML dashboard file (no external libraries, everything inline) that cleans this data and reports on it.

Cleaning rules:

- order_id must be unique: if an id appears twice, keep the first row and drop the rest.
- Rows whose amount is not a number (like N/A or unknown) are excluded from revenue, but report how many you excluded.
- Negative amounts are refunds and must count as negative revenue.
- Region names must be normalized (TX, tx, Texas, texas are all Texas; CA is California; OH is Ohio).
- Dates appear in multiple formats; parse them all. Q1 = Jan-Mar, Q2 = Apr-Jun.

The dashboard must show:

1. A pivot table of total revenue by region (rows) and quarter (columns), with row totals, column totals, and a grand total, formatted as dollars with two decimals.
2. A bar chart of revenue by region drawn with inline SVG or canvas (no chart libraries).
3. A data quality panel: rows received, duplicates dropped, rows excluded, rows counted.
4. A one-line headline stating the top region and its total.

The numbers displayed must be computed correctly from the data. Output only the complete HTML file.

```
order_id,order_date,region,product,amount_usd
1001,2025-01-06,Texas,widget,"1,240.00"
1002,1/9/2025,CA,gadget,$980
1003,2025-01-14,TX,gizmo,312.50
1004,"Jan 21, 2025",california,widget,"1,764.00"
1005,2025-01-27,Ohio,gadget,455
1006,2/3/2025,texas ,widget,N/A
1007,2025-02-10,OH,gizmo,"1,102.25"
1008,"Feb 14, 2025",California,widget,"2,310.00"
1009,2025-02-18,TX,gadget,-450.00
1010,2025-02-24,Texas,gizmo,870.10
1011,3/3/2025,Ohio,widget,"$1,999.99"
1012,2025-03-08,CA,gizmo,640.00
1008,"Feb 14, 2025",California,widget,"2,310.00"
1013,"Mar 12, 2025",tx,gadget,"1,580.00"
1014,2025-03-19,ohio,gizmo,225.75
1015,2025-03-27,California,gadget,"3,050.00"
1016,4/2/2025,TX,widget,990.00
1017,2025-04-07,Ohio,gadget,unknown
1018,"Apr 11, 2025",CA,widget,"1,420.50"
1019,2025-04-16,Texas,gizmo,"2,075.00"
1020,2025-04-22,OH,widget,$760.25
1021,5/1/2025,California,gizmo,"1,111.11"
1022,2025-05-06,TX,gadget,845.00
1023,"May 13, 2025",Ohio,gizmo,-220.00
1024,2025-05-19,texas,widget,"1,300.00"
1025,2025-05-28,CA,gadget,"2,940.40"
1026,6/4/2025,Ohio,widget,515.00
1027,2025-06-10,California,widget,"1,875.25"
1028,"Jun 17, 2025",TX,gizmo,N/A
1029,2025-06-23,OH,gadget,"1,660.00"
1030,2025-06-30,Texas,gadget,730.90
```

Using only raw inline SVG code (no JavaScript, no external images, no libraries), draw this exact scene as a single self-contained HTML file:

A happy dog wearing a yellow hard hat is driving a forklift. The forklift is lifting a rack of three stacked computer servers on its front forks. The servers have small colored indicator lights. The scene is outdoors: flat ground, a sun in the top left, two clouds, and a small house far in the background on the right. The forklift has two wheels touching the ground and the dog sits in the driver seat holding the steering wheel.

Layering must be correct (wheels touch the ground, the dog is inside the forklift, the hat is on the dog's head, the servers sit on the forks). Sign the artwork by writing your own model name in small text in the bottom right corner.