Qwen3.5-2B 與 Gemma 4 E2B 入門級 GPU 評測：Qwen 不用投機解碼時較快，Gemma 答案質素較高

This article provides a step by step deployment guide for Qwen3.5-2B to a local, desktop and laptop hosted GPU enabled system, served by llama.cpp beside Gemma 4 E2B on the same cards. A suite of Python MCP tools is built to simplify management of the llama.cpp hosted deployment.

https://github.com/xbill9/qwen-dev/tree/main/local-llamacpp-3050oem-qwen352b-q4_0

Qwen3.5-2B decodes 1.12x as fast as Gemma 4 E2B on an RTX 3050 and prefills 1.24–1.31x as fast. On three tasks, with thinking off for both models, Gemma answers better on every one: its 4-bit build leads Qwen's 4-bit build by 15.5 to 26.0 points and Qwen's unquantized bf16 by 10.0 to 19.5. With each model's drafter turned on, the speed lead disappears. The gaps survive scoring every cut-off or tool-less Qwen answer as right.

What is this project trying to Do?

The rigs in this repository serve Gemma 4 E2B from local hardware: an RTX 3050 OEM in a desktop and a GTX 1650 Ti in a laptop. Both run Google's quantization-aware-trained (QAT) 4-bit build, and both have measured scores for it on the same three tasks.

The question here: what does a Qwen of the same size class give on the same cards? Qwen3.5-2B counts 1.943B parameters in its text GGUF. Everything below runs both models with the same flags and the same task code, and on one llama.cpp binary except where the Scope paragraph says otherwise.

At This Point You Should Have…

• An NVIDIA GPU with a CUDA driver — here an RTX 3050 OEM (8 GB, Ampere, driver 615.78.08) and a GTX 1650 Ti (4 GB, Turing)
• llama.cpp built with CUDA — commit 7182ef71e here
• Python 3 for the converter and the MCP server — the system python3, no virtualenv
• The repository cloned, and local-llamacpp-3050oem-qwen352b-q4_0/ as your working directory
• Claude Code, or any MCP client that speaks stdio

The Two Models

Only Qwen's 6 full-attention layers hold a KV cache. The 18 DeltaNet layers keep a fixed-size state per request, so Qwen's prefill cost grows more slowly with prompt length.

Step 1 — Build the Qwen Q4_0 File

Qwen publishes bf16 weights only, so the rig builds its own 4-bit file. make quantize converts the Hugging Face checkout to a bf16 GGUF and runs the standard llama-quantize Q4_0, with no importance matrix:

python3 ~/llama.cpp/convert_hf_to_gguf.py ~/models/Qwen3.5-2B --outtype bf16 --outfile Qwen3.5-2B-bf16.gguf
~/llama.cpp/build/bin/llama-quantize Qwen3.5-2B-bf16.gguf Qwen3.5-2B-Q4_0.gguf Q4_0

The result, from the rig's tpu.env:

#   convert_hf_to_gguf.py --outtype bf16   -> 3,897,388,000 B, 335 tensors
#   llama-quantize ... Q4_0 (no imatrix)   -> 1,236,741,088 B
# llama-quantize's standard Q4_0 mix: 194 matrices Q4_0 (806.4 MB), token_embd
# Q6_K (417.2 MB, which is also the output head since it is tied), 140 F32 norms

This file is plain rounding from bf16. Gemma's Q4_0 file holds weights Google trained to sit on the 4-bit grid. Step 7 measures what that difference is worth.

Step 2 — Serve It

make serve runs llama-server with the Gemma rig's flags, so the two serve the same 131,072-token window:

llama-server -m Qwen3.5-2B-Q4_0.gguf --host 0.0.0.0 --port 8080 -ngl 99 -c 131072 \
  -ctk f16 -ctv f16 -fa 1 -t 6 -tb 12 --parallel 1 --reasoning off --metrics

Qwen3.5 thinks by default. --reasoning off puts answers in content, as on the Gemma rig; a request can still turn thinking on with "chat_template_kwargs": {"enable_thinking": true}.

At the full context it uses 2995 MiB of the 3050's 8192, and 2951 MiB of the 1650 Ti's 4096. The server log shows the MTP layer loaded and left idle:

0.01.266.683 W model has unused tensor blk.24.attn_norm.weight (size = 8192 bytes) -- ignoring
0.01.266.690 W model has unused tensor blk.24.post_attention_norm.weight (size = 8192 bytes) -- ignoring
0.01.266.699 W model has unused tensor blk.24.attn_q.weight (size = 4718592 bytes) -- ignoring

Step 3 — Check That llama.cpp Runs Qwen Correctly

A low score could come from the runtime. Three checks cover the prompt going in, the probabilities coming out and the tool calls Qwen writes.

Prompts. Every prompt was rendered through llama-server's /apply-template and through transformers' apply_chat_template. All 400 tool-calling prompts, with their tools, and 100 math prompts came out byte-identical.

Next-token probabilities. The reference is transformers in float32 on the CPU over Qwen's bf16 weights, on 65 prompts from the quality suite:

llama.cpp's bf16 file sits closer to the float32 reference than transformers' own bf16 path does. The drift comes from the 4-bit rounding.

Tool calls. Every BFCL reply that made no tool call was searched for text that looks like a call the parser could have missed (tool_call, <function, a JSON object): 0 of 79 at Q4_0 and 0 of 48 at bf16. Those replies solve the problem in prose, such as working the Pythagorean theorem step by step.

Step 4 — Measure Speed on One Binary

llama-bench ran both models on the same build, card and hour, five cells alternating between models, each started with the GPU at 40 °C or below:

llama-bench -p 512,2048 -n 128 -r 5 -ngl 99 -fa 1 -t 6

The summary script's output on the RTX 3050:

pp512: qwen 5193.44  gemma 4203.68  qwen/gemma 1.235x  qwen fa1/fa0 1.042x  qwen rep/first 0.993x  gemma rep/first 0.997x
pp2048: qwen 5316.13  gemma 4050.45  qwen/gemma 1.312x  qwen fa1/fa0 1.078x  qwen rep/first 0.996x  gemma rep/first 0.999x
tg128: qwen 132.84  gemma 118.30  qwen/gemma 1.123x  qwen fa1/fa0 1.012x  qwen rep/first 1.001x  gemma rep/first 1.002x

Qwen's prefill rises with prompt length (5193 to 5316 t/s) while Gemma's falls (4204 to 4050). That fits the DeltaNet layers, whose cost does not grow with the prompt.

Step 5 — Measure Answer Quality

Three tasks, the same task code and settings as the Gemma runs, 8 requests at a time, greedy:

• Suite: 13 classification sets, 3,880 questions, scored from the next-token probabilities
• GSM8K: 1,319 math word problems, 2,048-token limit
• BFCL simple: 400 tool-calling requests

Qwen's suite prompts end in Qwen's own chat template with thinking closed, and its answers are read from Qwen's own tokens. Every pair of runs is compared question by question, with a 95% range from resampling.

Every range is far from zero. On the suite, Qwen is ahead only on boolq and squad2.

Where Qwen loses points:

• Option letters. Asked for A, B or C, it often writes the option's name instead. On multinli it answers neutral on 298 of 299 questions.
• One-sided answers. It calls 248 of 300 comments toxic, where the gold labels mark 32.
• Long math answers. 100 GSM8K replies hit the 2,048-token limit, against 5 for Gemma. With thinking off it still deliberates inside the answer.
• Prose instead of tool calls. 79 of its 116 wrong BFCL answers made no tool call, mostly by solving the problem in text. Gemma skipped the tool 6 times.

Step 6 — Score Qwen Unquantized

Qwen's 4-bit file is plain rounding, so part of the gap may be the rounding. The bf16 GGUF from Step 1 ran through the same three tasks on the same binary, the same day:

• Rounding to Q4_0 costs Qwen 5.5 to 6.5 points per task, every range clear of zero.
• Gemma's 4-bit build still leads unquantized Qwen by 19.5, 10.0 and 16.0 points. The rounding accounts for a quarter to a third of the gap; the rest is the model.
• The same habits show at bf16: most comments called toxic, 48 BFCL replies with no tool call, 52 GSM8K replies at the length limit.

Step 7 — Measure What QAT Is Worth to Gemma

Gemma's 4-bit file comes from quantization-aware training; Qwen's comes from rounding. To compare like with like, Gemma E2B's ordinary bf16 release was rounded to Q4_0 with the exact recipe used for Qwen, and scored on the GTX 1650 Ti beside the other three:

QAT is worth 2 to 5 points to Gemma at 4 bits. Rounded Gemma still beats unquantized Qwen on all three tasks, by 14.4, 7.4 and 13.3 points.

Is the Comparison Fair to Qwen?

Four objections a Qwen user will raise, each checked against the per-question results.

Gemma is the bigger model. llama-bench counts 4.629B parameters for Gemma E2B against 1.943B for Qwen, most of Gemma's in per-layer embeddings. By file size, Qwen's bf16 GGUF is 3.90 GB and Gemma's QAT Q4_0 is 2.61 GB, and the larger Qwen file loses every task to the smaller Gemma one.

The token limit cut Qwen's math answers off. Scoring every GSM8K reply that hit the 2,048-token limit as right:

The tool-calling task punishes prose answers. Scoring every BFCL reply with no tool call as right:

The suite scores answer format. Partly right. On three subsets Qwen gives one answer almost every time or writes the option name in place of the letter: civil_comments, summeval-consistency and multinli. With those removed, the bf16 gap to Gemma on the GTX 1650 Ti shrinks from 19.3 points to 11.2. The suite is the most format-sensitive of the three results; Gemma still leads on it.

Q4_K_M and importance-matrix builds of Qwen were not scored. The bf16 file is the most any 4-bit build can recover, and Gemma leads it on every task.

Step 8 — Run the Bf16 File on a 4 GB Card

Qwen's bf16 GGUF holds 3.62 GiB of weights, and the 1650 Ti has 4 GiB. With every layer on the card, the 501.52 MiB compute buffer fails to allocate. Moving the last 6 of 24 blocks to the CPU is the fewest that fits:

llama-server -m Qwen3.5-2B-bf16.gguf -ngl 99 -ot 'blk\.(18|19|20|21|22|23)\.=CPU' \
  -c 8192 -ctk f16 -ctv f16 -fa 1 -t 6 -tb 12 --parallel 1 --reasoning off --metrics

Decode drops to 0.24x and prefill holds at 0.96x. The split costs no score: bf16 on this card scored 0.488, 0.807 and 0.762, level with the 3050.

Step 9 — Draft With Qwen's MTP Head

The Q4_0 file already carries Qwen's multi-token-prediction layer. llama.cpp drafts from it with --spec-type draft-mtp and no extra file:

llama-server -m Qwen3.5-2B-Q4_0.gguf -ngl 99 -c 8192 -fa 1 \
  --spec-type draft-mtp --spec-draft-n-max 1

The log confirms it and reports acceptance per request:

0.02.045.088 I common_speculative_init_result: creating MTP draft context against the target model '/home/xbill/models/Qwen3.5-2B-gguf/Qwen3.5-2B-Q4_0.gguf'
0.04.830.818 I slot print_timing: id  0 | task 8 | draft acceptance = 0.91729 (  122 accepted /   133 generated), mean len =  1.92

Six prompts (code, chat, story, explain, facts, summary), one request at a time, geometric mean against no drafting on the RTX 3050:

• 1 drafted token is the setting to serve. Code gains most (1.25x greedy); story least (1.09x).
• Greedy replies stay byte-identical to no drafting up to 3 drafted tokens on the 3050.
• On the same card, Gemma's separate drafter gave 1.39x at 2 tokens.

Step 10 — Race the Two Drafters on One Card

On the GTX 1650 Ti, both models on one binary, one request at a time, each with its best drafter setting of 1 token:

Sampled at temperature 0.8 the result is the same: 86.1 against 90.7 tok/s. Gemma's drafter gains more, so the two tie.

Drafting leaves Qwen's answers level: GSM8K 0.748 and BFCL 0.698 with the MTP head, against 0.734 and 0.688 without. At 8 requests at once, each task took 1.08–1.09x as long with drafting.

🔎 Tip: Pass Qwen's Tokenizer to Probability-Scored Tasks

A task that scores answers from next-token probabilities has to look up the answer labels in the model's own vocabulary. Qwen's differs from Gemma's, so give the scorer --tokenizer Qwen/Qwen3.5-2B; the default Gemma ids would read the wrong tokens.

🔎 Tip: Two Hosts, Two Different Hashes

The 3050 host and the 1650 Ti host built Qwen's Q4_0 from the same revision on the same llama.cpp commit, and got files of the same size with different SHA-256s (1c27608b… and bd99d5ae…). The bf16 files differ too, so the change enters at conversion. They pick the same answer on 62 of 65 prompts and score within 2.2 points on every task. Pin a hash per host, and keep one file if you need byte-identical results across machines.

Compare and Contrast

So, Which One?

Gemma 4 E2B. Qwen3.5-2B decodes faster without a drafter and prefills faster on prompts up to 2,048 tokens. Gemma answers better on all three tasks measured, at 4 bits against Qwen's bf16, and with both drafters on the speed difference is gone.

Qwen's built-in MTP head is a convenience: drafting with no extra file, and greedy text unchanged at 1 drafted token. Choose Qwen when prefill speed matters more than the answers; how its lead grows past 2,048-token prompts is unmeasured here.

Summary

The goal of this article was to serve Qwen3.5-2B with llama.cpp beside Gemma 4 E2B on two budget GPUs and compare them on speed and answer quality. The key to the solution was running both models on one llama.cpp binary with the same flags and the same task code, and scoring Qwen at bf16 and Gemma rounded the way Qwen's file is built. The results were:

• 🟢 llama.cpp runs Qwen3.5-2B correctly: identical prompts to Qwen's template, and its bf16 probabilities sit closer to float32 than transformers' bf16 does.
• 🟢 Qwen is faster without drafting: 1.12x decode and 1.24–1.31x prefill on the RTX 3050, 1.05x and 1.43–1.51x on the GTX 1650 Ti.
• ❌ Qwen scores lower on all three tasks: 15.5 to 26.0 points behind Gemma at 4 bits, 10.0 to 19.5 behind at bf16.
• 🟢 The gaps hold in Qwen's favour: scoring every cut-off GSM8K reply and every tool-less BFCL reply as right still leaves Qwen behind.
• ⚠️ Rounding is a quarter to a third of Qwen's gap, and QAT is a tenth to a fifth of Gemma's lead.
• 🟢 Qwen's MTP head drafts with no extra file: 1.15x greedy at 1 token, greedy text unchanged.
• ⚠️ With both drafters on, the two tie on speed: 91.0 against 92.1 tok/s on the GTX 1650 Ti.

Scope: one RTX 3050 OEM (8 GB) and one GTX 1650 Ti (4 GB), llama.cpp 7182ef71e for every Qwen run and for Gemma on the 1650 Ti, measured 2026-10-16 and 2026-10-17. The Gemma quality figures on the 3050 come from 2026-10-13 on llama.cpp 77c4540ce; Gemma's runs on the 1650 Ti on 7182ef71e scored level with them. Speed cells ran 3 to 5 repeats; each quality task ran once per file. Both models ran with thinking off and greedy decoding throughout; Qwen thinks by default, and thinking on was not run for either. The three tasks are classification, grade-school math and simple tool calling: no coding, Chinese, instruction-following or long-context task. Prefill was measured up to 2,048 tokens and every Qwen run served one request at a time, except the 8-request quality runs. No importance-matrix, Q4_K_M or Q8_0 build of either model was scored. The scoring code and the two cards were set up for this repository's Gemma work.

The strategy for using MCP for serving and comparing Qwen3.5-2B and Gemma 4 E2B on local GPUs was validated with an incremental step by step approach.

References

• local-llamacpp-3050oem-qwen352b-q4_0 | GitHub
• local-llamacpp-1650ti-qwen352b-q4_0 | GitHub
• Qwen/Qwen3.5-2B | Hugging Face
• xbill9/gemma-4-E2B-it-qat-q4_0-exact-gguf | Hugging Face
• google/gemma-4-E2B-it | Hugging Face
• Gemma 4's MTP Drafter, Explained | dev.to
• llama.cpp | GitHub