Qwen3.8 開源發布#
Qwen團隊在8月13日開源了Qwen3.8-2.4T-A95B,對標網頁服務的Qwen3.8-MAX,這也是他們首個MAX等級的開源模型。先前最大的只有Qwen3.5-397B-A17B,光是這個參數量的提升就能看出來,Qwen3.8的開源可以說是來勢洶洶,左打Kimi-K3,右踢GLM-5.2,非常想奪回開源模型界的TOP寶座。從2.4T-A95B的Model Card來看,在大部份能力方面,可以說是非常接近各家的主流旗艦大模型。
但是呢,這種Trillion級別的大模型,應該說…「超大」模型,本來不是我們這種一般人目光所及的。2.4T的參數8-bit(FP8)量化都還要2.5TB左右的VRAM,就算壓到4-bit(NVFP4)都至少要1.5TB的VRAM,這還是沒算上metadata與推論額外開銷的粗略計算。誰家裡會有1.5TB的VRAM啊? 四臺Mac Studio 512GB嗎…
所以我的目光就放在小尺寸模型的消息,聽說Qwen團隊要釋出27B Dense模型就很興奮,今天Qwen3.8 27B來了!我終於有機會換Local LLM Model啦!目前在公司內部使用Qwen3.6 35B-A3B也沒什麼大問題,畢竟大家都知道地端模型的能力極限就在那,日常的高強度推理或開發還是會交給外面的大模型,如GPT-5.6 Sol或Claude Opus 5等來執行,地端模型主要還是公司內的一些簡單Agent操作以及知識庫問答的部分,但是Qwen3.6畢竟還是有一點點舊了,各項小BUG累積起來讓我想找個機會換掉他,這個機會來啦!
自從得知Qwen的前朝元老離開後,一直很擔心Qwen從此走向閉源,之後的Qwen3.7就是完全沒開源,一度放棄了Qwen想去尋找其他參數量相近的模型。沒想到Qwen3.8驚喜開源,還能看到Qwen團隊持續在開源模型,希望他們能保持繼續做開源模型,尤其是小尺寸(~40B)的開源模型,這個大小的模型最適合跑在消費級硬體上,有助於開源社群的發展。

模型選擇#
不是早就決定好模型是Qwen 3.8 27B了嗎? 還要選什麼? 除了模型以外,還有「量化」要選擇,更重要的是「誰做的量化?」,我會偏向於選擇自己信任的來源,至少要是一個看起來沒什麼大問題的單位提供的模型。
Qwen3.8 27B BF16模型權重大小約55.6GB實在是太大了,DGX-Spark對NVFP4的模型是有最佳化的,因此我一開始就往NVFP4去找。過往在Qwen3.6與Nemotron3,我會選擇來自nvidia或RedHatAI量化的NVFP4模型,但是這次Qwen3.8發布首日,這兩家都還沒發布。
在HuggingFace搜索一番後,發現一間名為Inferact Inc.的公司有量化Qwen3.8 27B NVFP4,深入研究後發現這Inferact公司不得了,他是vllm的核心成員們創立的公司,公司裡面那幾個人就是vllm論文上面的作者群,甚至第一作者就在裡面。vLLM Recipes網頁上面也有對於Qwen3.8 27B的Day0支援,recipe上面的NVFP4就是使用來自Inferact量化的模型。更重要的是,Inferact/Qwen3.8-27B這個Repo內的檔案幾乎都是在Qwen3.8公布前就已經上傳,搭配vLLM官方的Day-0支援資訊來看,這個模型的來源可靠性應該沒有大問題,本次測試就用這個量化模型來跑。
recipes之前只是一個簡陋的Markdown mkdocs網站,各種參數跟use case塞在一起,較難取得完整資訊,UX體驗非常不好。自從前陣子改版之後,變成非常專業好用,硬體跟量化變體,各項參數只要點一點就能出vllm指令,複製回自己的recipe稍微改一下就好,幾乎做到開箱即用
https://recipes.vllm.ai/Qwen/Qwen3.8-27B?variant=nvfp4
應該不用多久unsloth也會發布llama.cpp適用的GGUF格式,然後各路民間高手會開始拿頂級大模型拷打(X)微調(O) Qwen3.8 27B,沒幾天就會有各種後墜長到不行的衍伸型號出現。
測試環境#
本次測試環境依然是NVIDIA DGX Spark,也就是上次文章 DGX-Spark LLM模型選擇 (2026Q2) 中的兩臺,系統更新到最新,vLLM image也不同,詳細系統版本如下
- Linux Kernel: 6.17.0-1029-nvidia
- Driver Version: 580.173.02
- CUDA Version: 13.0
- Docker Version : 29.2.1
- nvcc Version: 13.0.88
vLLM透過vllm/vllm-openai docker image執行,vLLM Recipes明確說明Qwen3.8需要 vLLM v0.27.2 以上的版本,測試時這個版本還沒正式build到docker hub上面,因此本次測試使用的是 vllm/vllm-openai:qwen38 ,看image tag應該是給Qwen3.8 Day0支援用的image,就先選他跑看看。
目前vLLM Recipe對Qwen3.8的版本標記改為v0.17.0+,我的環境都是0.27.1以上了,就沒測試這部分
參數設定#
vLLM Recipes官方給的啟動參數如下
vllm serve Inferact/Qwen3.8-27B-NVFP4 \
--tensor-parallel-size 1 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--reasoning-parser qwen3 \
--mm-encoder-tp-mode data稍做修改,貼入我的docker compose yaml文件裡面
這份yaml非最終版本,會隨著測試會增加不同參數,先不要抄這份作業,下面有完整版
| |
- 我自己習慣另外下載模型到指定路徑,而不是啟動時傳入HF_TOKEN讓他自動下載,所以啟動參數會直接指向模型所在資料夾
gpu-memory-utilization 0.5027B+NVFP4用完DGX-Spark所有的Unified Memory太浪費了,只給他用0.5左右,剩餘的RAM拿去做其他事情,最大化利用效率
測試#
速度測試#
速度方面的測試同樣使用 llama-benchy Version 0.3.7
與上次在DGX-Spark LLM模型選擇 (2026Q2) 文章中使用的相同版本
其實只是我懶得更新而已
原生速度
就是上面那一份yaml直接跑起來的速度
| model | test | t/s (total) | t/s (req) | peak t/s | peak t/s (req) | ttfr (ms) | est_ppt (ms) | e2e_ttft (ms) |
|:--------|------------:|----------------:|-----------------:|-------------:|-----------------:|------------------:|------------------:|------------------:|
| qwen3.8 | pp2048 (c1) | 2264.18 ± 52.89 | 2264.18 ± 52.89 | | | 909.55 ± 21.51 | 905.46 ± 21.51 | 909.55 ± 21.51 |
| qwen3.8 | tg128 (c1) | 9.88 ± 0.01 | 9.88 ± 0.01 | 10.00 ± 0.00 | 10.00 ± 0.00 | | | |
| qwen3.8 | pp2048 (c2) | 2317.67 ± 1.94 | 1333.45 ± 171.97 | | | 1566.80 ± 201.42 | 1562.71 ± 201.42 | 1566.80 ± 201.42 |
| qwen3.8 | tg128 (c2) | 18.92 ± 0.01 | 9.68 ± 0.07 | 20.00 ± 0.00 | 10.00 ± 0.00 | | | |
| qwen3.8 | pp2048 (c4) | 2312.70 ± 2.84 | 832.05 ± 302.18 | | | 2735.87 ± 754.28 | 2731.79 ± 754.28 | 2735.87 ± 754.28 |
| qwen3.8 | tg128 (c4) | 32.66 ± 0.03 | 9.03 ± 0.36 | 40.00 ± 0.00 | 10.00 ± 0.00 | | | |
| qwen3.8 | pp2048 (c8) | 2298.21 ± 0.37 | 560.64 ± 331.09 | | | 4664.26 ± 1886.44 | 4660.17 ± 1886.44 | 4664.26 ± 1886.44 |
| qwen3.8 | tg128 (c8) | 50.89 ± 0.01 | 7.88 ± 0.72 | 72.00 ± 0.00 | 9.50 ± 0.50 | | | |開啟MTP=3之後速度
加入以下參數
--speculative-config '{"method":"mtp","num_speculative_tokens":3}'
| model | test | t/s (total) | t/s (req) | peak t/s | peak t/s (req) | ttfr (ms) | est_ppt (ms) | e2e_ttft (ms) |
|:--------|------------:|-----------------:|-----------------:|--------------:|-----------------:|------------------:|------------------:|------------------:|
| qwen3.8 | pp2048 (c1) | 1854.56 ± 87.22 | 1854.56 ± 87.22 | | | 1110.72 ± 50.40 | 1107.03 ± 50.40 | 1110.72 ± 50.40 |
| qwen3.8 | tg128 (c1) | 14.69 ± 0.35 | 14.69 ± 0.35 | 22.33 ± 1.25 | 22.33 ± 1.25 | | | |
| qwen3.8 | pp2048 (c2) | 1867.05 ± 194.87 | 1170.74 ± 355.40 | | | 1905.61 ± 506.51 | 1901.92 ± 506.51 | 1905.61 ± 506.51 |
| qwen3.8 | tg128 (c2) | 25.95 ± 1.27 | 14.30 ± 1.31 | 39.33 ± 2.36 | 22.33 ± 2.49 | | | |
| qwen3.8 | pp2048 (c4) | 1723.87 ± 398.68 | 757.83 ± 372.44 | | | 3447.07 ± 1716.67 | 3443.38 ± 1716.67 | 3447.07 ± 1716.67 |
| qwen3.8 | tg128 (c4) | 37.60 ± 3.82 | 12.15 ± 1.36 | 69.00 ± 1.41 | 20.75 ± 2.45 | | | |
| qwen3.8 | pp2048 (c8) | 2003.38 ± 4.49 | 525.81 ± 316.08 | | | 5091.69 ± 2211.95 | 5088.00 ± 2211.95 | 5091.69 ± 2211.95 |
| qwen3.8 | tg128 (c8) | 58.36 ± 2.91 | 10.47 ± 1.36 | 120.33 ± 6.18 | 19.38 ± 1.55 | | | |開啟MTP=3與 KV=FP8之後速度
加入以下參數
--speculative-config '{"method":"mtp","num_speculative_tokens":3}'--kv-cache-dtype fp8
| model | test | t/s (total) | t/s (req) | peak t/s | peak t/s (req) | ttfr (ms) | est_ppt (ms) | e2e_ttft (ms) |
|:--------|------------:|----------------:|-----------------:|--------------:|-----------------:|------------------:|------------------:|------------------:|
| qwen3.8 | pp2048 (c1) | 1826.55 ± 73.79 | 1826.55 ± 73.79 | | | 1127.13 ± 44.52 | 1123.59 ± 44.52 | 1127.13 ± 44.52 |
| qwen3.8 | tg128 (c1) | 14.34 ± 0.98 | 14.34 ± 0.98 | 21.33 ± 1.70 | 21.33 ± 1.70 | | | |
| qwen3.8 | pp2048 (c2) | 1903.44 ± 11.24 | 1073.55 ± 120.63 | | | 1936.67 ± 216.78 | 1933.13 ± 216.78 | 1936.67 ± 216.78 |
| qwen3.8 | tg128 (c2) | 28.34 ± 2.65 | 15.53 ± 1.93 | 40.00 ± 2.94 | 21.50 ± 1.89 | | | |
| qwen3.8 | pp2048 (c4) | 1962.88 ± 2.86 | 746.37 ± 273.82 | | | 3086.95 ± 945.79 | 3083.40 ± 945.79 | 3086.95 ± 945.79 |
| qwen3.8 | tg128 (c4) | 42.13 ± 1.85 | 13.55 ± 1.83 | 71.67 ± 3.30 | 21.00 ± 2.35 | | | |
| qwen3.8 | pp2048 (c8) | 1957.76 ± 1.56 | 509.40 ± 300.57 | | | 5228.59 ± 2254.72 | 5225.04 ± 2254.72 | 5228.59 ± 2254.72 |
| qwen3.8 | tg128 (c8) | 60.48 ± 3.13 | 10.82 ± 1.87 | 122.33 ± 3.09 | 19.08 ± 2.04 | | | |開啟MTP=3 與 KV=FP8 與 max-num 之後速度
加入以下參數
--speculative-config '{"method":"mtp","num_speculative_tokens":3}'--kv-cache-dtype fp8--max-num-seqs 8--max-num-batched-tokens 8192
| model | test | t/s (total) | t/s (req) | peak t/s | peak t/s (req) | ttfr (ms) | est_ppt (ms) | e2e_ttft (ms) |
|:--------|------------:|-----------------:|----------------:|--------------:|-----------------:|------------------:|------------------:|------------------:|
| qwen3.8 | pp2048 (c1) | 1907.53 ± 23.85 | 1907.53 ± 23.85 | | | 1078.10 ± 13.80 | 1073.99 ± 13.80 | 1078.10 ± 13.80 |
| qwen3.8 | tg128 (c1) | 15.04 ± 1.42 | 15.04 ± 1.42 | 21.33 ± 0.94 | 21.33 ± 0.94 | | | |
| qwen3.8 | pp2048 (c2) | 1912.38 ± 255.40 | 988.89 ± 100.44 | | | 2101.32 ± 250.13 | 2097.21 ± 250.13 | 2101.32 ± 250.13 |
| qwen3.8 | tg128 (c2) | 25.52 ± 2.13 | 13.30 ± 1.04 | 38.00 ± 3.56 | 20.83 ± 2.19 | | | |
| qwen3.8 | pp2048 (c4) | 2119.65 ± 52.39 | 540.76 ± 20.14 | | | 3798.37 ± 142.34 | 3794.25 ± 142.34 | 3798.37 ± 142.34 |
| qwen3.8 | tg128 (c4) | 51.59 ± 1.31 | 13.73 ± 0.83 | 71.33 ± 2.36 | 21.25 ± 1.23 | | | |
| qwen3.8 | pp2048 (c8) | 2109.32 ± 1.69 | 289.94 ± 61.69 | | | 7293.65 ± 1048.76 | 7289.53 ± 1048.76 | 7293.65 ± 1048.76 |
| qwen3.8 | tg128 (c8) | 70.84 ± 1.06 | 12.79 ± 1.64 | 127.00 ± 1.63 | 19.88 ± 1.69 | | | |參數修正#
結合上面llama-benchy的測試結果,接下來的測試都會用這組參數去跑
| |
圖像測試#
將模型接上OpenWebUI(直接連接vLLM),然後在WebUI中上傳圖像與Prompt給模型,參數全預設,沒有做另外的調整,System Prompt為空
礙於版面的原因,我這邊只做一張圖片測試,模型的思考過程與輸出另外放在檔案中,如果想看詳細過程的再去點就好
# User Prompt
請描述圖片中的場景,並詳細解釋圖中發生的事情,以及選擇接下來的操作,並推測會發生的事情
思考用時3分10秒
展開/收折模型思考過程
The user wants me to describe the scene in the image, explain what's happening in detail, suggest the next action, and predict what will happen.
Let me analyze the image carefully.
This is a Pokémon game screenshot (Pokémon Legends: Arceus style UI, but the interface with PP counts and Tera/太晶化 suggests Pokémon Violet/Scarlet — actually the "太晶化" Tera feature and the battle UI with move list on the right, HP bars, the Y button for 招式說明 (move description) — this is Pokémon Scarlet/Violet).
Let me read the details:
Left side (player's Pokémon): 洛托姆 (Rotom), Lv.100, HP 304/304 (full). It's shown with a cyan/ice aura — wait, the aura color. Rotom is normally Electric type. The cyan/blue outline might indicate something. Actually Rotom's form here — it looks like the Rotom with a blue/teal glow which could be its Tera type or the ice move. Actually the move list shows 水炮 (Hydro Pump), 求雨 (Rain Dance), so Rotom is an Electric/??? type using water moves. The cyan outline around Rotom might just be its selected form or the Tera crystal.
Right side (opponent): 龍頭地鼠 (Armaidon), Lv.68, with a full HP bar (green). Armaidon is a Steel/Dark type Pokémon (the evolved form of Diglett/Dugtrio/Armaidon line — Armaidon is Steel/Dark).
Bottom center: 太晶化 (Tera) indicator with a green crystal and "R" button — this is the Tera button (press R to Tera Shift). The green color suggests the Tera type is... green = Grass? Actually the Tera crystal color here is greenish-yellow. Hmm.
Move list (right side):
十萬伏特 (Thunderbolt) - Electric - ×沒有效果 (no effect) - 15/15 PP
水炮 (Hydro Pump) - Water - ◯效果絕佳 (super effective) - 5/5 PP
伏特替換 (Volt Switch) - Electric - ×沒有效果 (no effect) - 20/20 PP
求雨 (Rain Dance) - Water - 5/5 PP
Now the analysis:
The scene: A battle between the player's Lv.100 Rotom (full HP 304/304) and the opponent's Lv.68 Armaidon (full HP).
Why is Thunderbolt "no effect"? Armaidon is Steel/Dark type. Steel is immune to Electric (and Ground, Electric). So Thunderbolt (Electric) has no effect because of Steel immunity. Volt Switch (Electric) also no effect for the same reason.
Hydro Pump (Water) is "super effective" (效果絕佳). Armaidon is Steel/Dark. Water is super effective against Steel. So Hydro Pump is super effective.
So the best move is clearly Hydro Pump (水炮), which is super effective.
But wait — there's also the Tera option. The green Tera crystal. If Rotom Terastalls, what type? The aura around Rotom is cyan/teal which could be a Tera type. Hmm. Actually if Rotom's Tera type is Water, then Thunderbolt would still be no effect (Electric move, Steel immune). Tera doesn't change move typing.
Actually let me think about whether to use Rain Dance (求雨) first to boost Water move power, then Hydro Pump. Rain Dance boosts Water moves by 50%. That's a strategic option. But Hydro Pump is super effective and Rotom is Lv.100 vs Lv.68, so Hydro Pump likely one-shots Armaidon anyway.
The question asks: describe the scene, explain what's happening, suggest the next action, and predict what will happen.
Let me also note: the player is in a higher-level situation (Lv.100 vs Lv.68). This looks like it could be a battle in the main story or post-game, perhaps against a team member or in the post-game. Actually the setting (dark cave-like background) suggests the Teal Mask / some cave, or the post-game "Battle Frontier" / "Post-game" area. Actually the dark rocky background could be the "Cascarra Cave" or the post-game "Blueberry Academy" or a cave in the main area. Hard to say. The dark cave background with purple tint is reminiscent of the "Slippage Well" or "Cascarra Cave" or the "Almond Town" area, or the post-game "Postgame" region. Actually it might be the "Cascarra Cave" in the main story where you battle Klemme, or it could be the post-game.
Hmm, actually the dark purple cave could be the "Slippage Well" where you fight Koa, or the "Cascarra Cave". But the level 100 suggests post-game. The dark cave could be the post-game "Cascarra Cave" or the "Teal Mask" related. Honestly I'll just describe it as a cave/underground battle.
Let me focus on the key gameplay logic:
Scene description: Pokémon Scarlet/Violet battle. Player's Rotom (Lv.100, full HP) facing opponent's Armaidon (Lv.68, full HP) in a dark cave.
What's happening: A turn-based battle. The player is choosing a move for Rotom.
Key analysis:
Armaidon is Steel/Dark type.
Steel is immune to Electric → Thunderbolt and Volt Switch do no damage (沒有效果).
Water is super effective against Steel → Hydro Pump (效果絕佳).
Rain Dance would boost Water moves.
Recommended next action: Use 水炮 (Hydro Pump). It's super effective, and given the level difference (100 vs 68) and full HP, it will likely KO Armaidon in one hit. Alternatively, use 求雨 (Rain Dance) first to boost, but that's less efficient since Hydro Pump already likely KOs.
Actually, one nuance: should the player consider Tera (太晶化)? The green crystal. If Rotom's Tera type is something that makes a currently-"no effect" move work... no, Tera doesn't bypass immunity. Immunity is based on typing, not Tera. So Tera wouldn't help Thunderbolt. So no point in Tera here. Actually, Tera could change Rotom's own typing for the opponent's next attack, but since we're likely KO-ing Armaidon, no need.
Also note: Volt Switch is "no effect" so it can't switch out via the move (it does no damage). But the player could switch manually. However, switching is unnecessary since Rotom can win.
Prediction: Using Hydro Pump will deal super-effective damage, likely reducing Armaidon's HP significantly or KOs it. Given the large level gap, Hydro Pump probably KOs Armaidon outright (or at least drops it to very low HP). Armaidon, being lower level and full HP, might try an attack but will likely be fainted or left with low HP.
Let me also double check Armaidon's typing and Water effectiveness: Armaidon = Steel/Dark. Water → Steel: super effective (2x). Water → Dark: neutral (1x). Combined = 2x (super effective). Yes, "效果絕佳" = super effective. Correct.
Electric → Steel: immune (0x). Electric → Dark: neutral. Combined = 0x (no effect). Correct, "沒有效果".
Great.
Now let me also mention the PP: Hydro Pump 5/5 (max), so plenty.
Now, the cyan aura on Rotom: This might indicate Rotom is in a specific Rotom form (e.g., Rotom-Frost? No, that's Ice). The cyan/teal outline... Actually in Scarlet/Violet, the selected Pokémon sometimes has a glow. The teal color might be Rotom's Tera type (Water? no, water is blue). Hmm the crystal is green. The outline is cyan. Hard to interpret. I'll mention Rotom has a cyan/teal aura which might reflect its Tera or form, but it's not critical.
Actually, wait — the green Tera crystal at bottom with "太晶化" label and R button. The color of the Tera crystal appears green/yellow-green. In the game, the Tera crystal color matches the Tera type. Green = Grass? Or it could be a generic glowing. The outline around Rotom being cyan could be... Rotom's base form glow. I won't over-interpret.
Let me write the response in Traditional Chinese (the user wrote in Traditional Chinese).
I'll structure: 場景描述 (scene description), 詳細解釋 (detailed explanation), 建議操作 (recommended action), 預測結果 (prediction).
Let me make it detailed and accurate.
One more thing about the "R" 太晶化 button: pressing R would Tera Shift Rotom. Given the situation, Tera is not necessary. I'll note that.
Also I should mention the Y = 招式說明 (move description) hint at bottom.
Let me also consider the strategic depth: Should the player use Rain Dance first? Rain Dance + Hydro Pump would be very powerful. But Hydro Pump alone is likely a KO. If there's any doubt about KO, Rain Dance guarantees it. But it costs a turn and Armaidon could attack. Given Rotom is Lv.100 with 304 HP and Armaidon Lv.68, Hydro Pump super effective will almost certainly KO. So just use Hydro Pump.
Alternatively, the player might want to Tera to Water if not already, to boost... but Tera to Water only boosts Water moves by 50% AND the move is already super effective. But Tera costs the turn and changes type. Not needed.
I'll present the recommended action as Hydro Pump, with a note about Rain Dance as an alternative for guaranteed KO if HP were uncertain, and explain why not to use the Electric moves or Tera.
Let me write it up nicely.展開/收折模型回答
## 場景描述
這是《Pokémon 朱/紫》的一場對戰畫面,背景是昏暗的洞窟/岩壁場地(紫黑色岩石地面與岩壁)。畫面呈現轉盤式對戰:
- **我方(左)**:`洛托姆`(Rotom),**Lv.100**,HP `304/304`(滿血),身上包著一道青藍色光暈。
- **對手(右)**:`龍頭地鼠`(Armaidon),**Lv.68**,HP 槽也是滿的。
- **畫面中下方**:綠色「太晶化」水晶標誌,旁有 `R` 鍵提示(按下可觸發太晶化)。
- **右下角招式選單**(含 PP 與屬性效果判斷):
1. `十萬伏特`(電)—— **× 沒有效果**,15/15
2. `水炮`(水)—— **◯ 效果絕佳**,5/5
3. `伏特替換`(電)—— **× 沒有效果**,20/20
4. `求雨`(水)—— 5/5
- 底部有 `Y 招式說明` 提示。
## 圖中正在發生的事情
這是一場**屬性剋制非常明顯**的對戰。關鍵在於雙方的屬性:
- **龍頭地鼠是「鋼/惡」屬性。**
- **鋼屬性對「電」屬免疫** → 所以兩個電系招式 `十萬伏特` 與 `伏特替換` 都被系統標為「× 沒有效果」,打上去完全沒傷害。
- **水屬性對鋼屬是 2 倍(超效)**,對惡屬是普通倍率 → 綜合下來 `水炮` 被標為「◯ 效果絕佳」。
- 我方洛托姆是 **Lv.100、滿血**,對手龍頭地鼠只有 **Lv.68**,級數差距極大(32 級),且我方面對的是超效命中。
簡單說:這是「電系無效、水系超效」的局勢,玩家正在為洛托姆挑選這一回合的招式。
## 建議的接下來操作
**直接选用 `水炮`(Hydro Pump)。** 理由:
1. 它是唯一能造成傷害且「效果絕佳」的攻擊招式,電系兩個招式對龍頭地鼠完全無效,等於廢招。
2. 32 級的級數差 + 滿血 + 超效命中,`水炮` 幾乎可以**一擊擊倒**龍頭地鼠。
3. **不建議先 `求雨`**:雖然 `求雨` 能再提升水系威力 50%,但 `水炮` 本身已經極可能一擊 KO,多耗一回合還可能讓龍頭地鼠反擊,得不償失。
4. **不建議 `太晶化`(R)**:太晶化不會取消「屬性免疫」,電系依然無效;而我們本來就要一擊結束,沒有必要改變屬性去挨打。
5. 也不必用 `伏特替換` 換場——它對龍頭地鼠無效,而且我方沒有換人的必要。
> 若萬一 `水炮` 沒能一擊 KO(例如龍頭地鼠有極高防禦或攜帶減傷道具),才考慮先 `求雨` 加威力再補刀;但在正常情況下 `水炮` 直接收尾。
## 預測會發生的事情
- 使用 `水炮` 後,水屬性以「效果絕佳」命中龍頭地鼠,造成巨量傷害;**在 Lv.100 對 Lv.68 的差距下,極可能直接擊倒(HP 歸零,龍頭地鼠倒下)**,對戰結束、我方獲勝。
- 即便極少數情況未一擊 KO,龍頭地鼠剩血也會極低;此時再補一發 `水炮`(5/5 PP 充足)即可收尾。
- 若選擇電系招式或 `太晶化`,則本回合不會造成任何傷害,反而可能挨打,是明顯不利的選擇。
**一句話結論**:這一局的答案很明確——**捨棄無效的電系,直接 `水炮` 一擊收尾**。模型的回答中間感覺有點亂寫,但是因為文字識別能力還不錯,最後被他矇到結果
相同的圖片丟給Qwen3.6有機會觸發讓他無限想下去的BUG,目前Qwen3.8還沒測出這種BUG
程式開發測試#
把模型接入OpenCode(直接對接vLLM),在OpenCode中提出網頁開發要求
# User Prompt
我要一個網頁呈現八大行星的太陽系模型,使用者要可以自由拉動時間,觀察一年之中行星之間的相對位置,網頁技術不限於python或是nodejs,你自行決定所有細節Plan Mode思考約8分鐘左右,提出一份計畫,切到Action Mode之後我直接請他開始實做
Action Mode思考了13分鐘才開始輸出,輸出用了5分鐘左右,之後就啟動了網頁伺服器有了以下的畫面

網頁的交互完全沒問題,右上角寫的功能全部都會正確作動,拖曳也都正常
最驚訝的是,這個開發過程模型似乎沒有查外部資訊,看起來都是模型本身的知識寫出來的
而且 one shot,一次就達成上面的結果
過往的Qwen3.6-35B-A3B大概需要來回修改10次左右,才能達到差不多的結果
結論#
從速度測試來看,原生的Recipe加上MTP、FP8 KV Cache與Batch限制參數後,單發速度可以從原本的9.88 t/s拉到15.04 t/s,四條併發有51.59 t/s,八條併發則可以到70.84 t/s,提升還算滿有感的。雖然我曾經在上篇文章寫過,我個人的模型速度下限是20 token/s,Qwen3.8 27B是一個Dense模型,顯然是較難以達到我的速度,15 token/s真的在實際使用上也不會太慢,多併發場景也都保持不錯的速度,考量到他較上一代出色的能力,屬於可接受的範圍。
| 設定 | TG c1 | TG c4 | TG c8 | PP2048 c1 | TTFT c1 |
|---|---|---|---|---|---|
| 原生 | 9.88 | 32.66 | 50.89 | 2264 | 910 ms |
| MTP=3 | 14.69 | 37.60 | 58.36 | 1855 | 1111 ms |
| MTP=3 + FP8 KV | 14.34 | 42.13 | 60.48 | 1827 | 1127 ms |
| 最終設定 | 15.04 | 51.59 | 70.84 | 1908 | 1078 ms |
圖像與開發能力倒是給了一點驚喜,Qwen團隊在前幾天開源Qwen3.8 2.4T-A95B留了一手,這個超大模型是沒有視覺能力的,必須透過外掛或是其他手段才能實現視覺功能,而這次的Qwen3.8 27B則是原生就具有視覺功能,而且在思考過程與辨識度上面相較於上一代有不錯的提升
當然,這只是發布這幾天的初步測試,詳細使用起來怎麼樣,可能過一段時間的體驗之後才會有比較多想法,到時候有機會在寫文章。
睽違幾個月之後再次開源的Qwen3.8模型,沒有意外又成為開源Local LLM的當紅炸子雞,30B左右的應該是除了他以外沒有更好的選擇 (Gemma 4: 你找我嗎?)
個人還是會希望有更高參數量的MoE模型出現,例如總參數量60B~100B左右的MoE模型,應該會非常適合放在DGX-Spark上面做Local Private Agent,並且兼顧速度與模型基本能力。
以上,是本人在Qwen3.8 27B首日測試的一點廢話 (雖然首日測試被我拖了三天才寫完 XD)