Coding Horizon

The Qwen Speed Run That Changes Local AI Forever

Every figure, name and benchmark the video puts on screen, chased to a primary source.


The model

Qwen3.8-27B, Alibaba’s Qwen family. Open weights on Hugging Face.

Claim Value Source
Parameters 27B, dense Qwen/Qwen3.8-27B
Layers 64 Qwen/Qwen3.8-27B
Native context 262,144 tokens Qwen/Qwen3.8-27B
Extended context 1,000,000 with YaRN, supported by vLLM and SGLang Qwen/Qwen3.8-27B
Vision Image and video input, native vision language model Qwen/Qwen3.8-27B
Thinking mode On by default, with reasoning_effort at xhigh, medium, low Qwen/Qwen3.8-27B
Multi token prediction Trained with MTP, usable by compatible inference engines Qwen/Qwen3.8-27B
Attention 3:1 hybrid of Gated DeltaNet linear attention and gated full attention, 24 query / 4 KV heads, head dim 256 Qwen/Qwen3.8-27B
Vocabulary 248,320 Qwen/Qwen3.8-27B
Predecessor Qwen3.6-27B Qwen/Qwen3.6-27B
Official FP8 checkpoint Published separately Qwen/Qwen3.8-27B-FP8

Benchmarks

Vendor reported on the official model card, comparing Qwen3.8-27B against Qwen3.6-27B. Independent replication is not yet available for these, so they are a ceiling rather than a measurement.

Benchmark Qwen3.8-27B Qwen3.6-27B
SWE-bench Pro 61.7 53.5
QwenSWEBench 79.0 49.3
OSWorld-Verified 84.3 63.9
Terminal Bench 2.1 73.0 63.4
LiveCodeBench v6 90.3 83.9
GPQA Diamond 89.2 87.8
IFBench 79.5 69.1

Source: Qwen/Qwen3.8-27B model card

The video states SWE-bench Pro, QwenSWEBench and OSWorld-Verified. The other four rows are recorded here for context and are not put on screen.

The 91 tokens per second run

A published deployment and tuning writeup for Qwen3.8-27B-Uncensored-FP8 on two RTX 5090 cards. Every figure below is from that writeup.

Hardware and configuration

Setting Value
GPUs 2 x RTX 5090, 32 GB each
Engine vLLM
Tensor parallel size 2
Model FP8 quantized checkpoint
Served context 131,072 tokens, of 262,144 native
GPU memory utilization 0.96
Max batched tokens 8,192
Prefix caching Enabled, xxhash
Speculative method qwen3_5_mtp, one speculative token

Measured decode throughput, single request, temperature 0, 220 completion tokens

Configuration Tokens per second
Baseline, bf16 KV cache, no prefix caching, no MTP 56.7
Prefix caching only, bf16 KV cache ~56.7
FP8 KV cache 76.3
MTP speculative decoding 88 to 91
MTP and FP8 KV cache together 49.3

Multi token prediction detail

Metric Value
Draft token acceptance rate 73.8%
Mean acceptance length 1.74 tokens

KV cache budget

KV cache dtype Budget
bf16 336,719 tokens
FP8 634,971 tokens

Source: fentz26/Qwen3.8-27B-5090

This is the source of the video’s four load bearing speed figures: 56.7 baseline, 76.3 with FP8 KV cache, 88 to 91 with MTP, and 49 when the two are combined. The regression when both are enabled is stated explicitly in the writeup, not inferred.

Prefix caching not moving raw decode is also from that writeup: it changes the cost of repeated prefill, not the speed of generation.

What the weights actually weigh

Build Size Source
BF16, 18 safetensor shards 55.56 GB decimal Qwen3.8-27B quantization comparison
Official FP8 repository ~30.88 GB Qwen3.8-27B quantization comparison
Q8 GGUF 26.90 to 29.30 GiB as above
Q6_K GGUF 20.89 to 23.29 GiB as above
Q5_K GGUF 17.27 to 18.83 GiB as above
Q4_K_M GGUF 15.66 to 17.67 GiB as above
IQ3_S GGUF 12.89 GiB as above
IQ2_S GGUF 10.38 GiB as above

Two notes on how these line up with the video.

The narration says BF16 is “roughly fifty four gigabytes just for the weights”. The official shards total 55.56 GB decimal, which is 51.7 GiB. A 27B model at two bytes per parameter is 54 GB by arithmetic. The narration’s figure sits inside that band and is stated as an approximation, so it is treated as sound.

The narration says Q6 GGUF “lands around the twenty one gigabyte class”. Q6_K is 20.89 to 23.29 GiB, and GiB is the unit a GPU reports free memory in, so the bottom of that range is the twenty one gigabyte class exactly. Confirmed.

The narration says four bit builds reach “the mid teens”. Q4_K_M is 15.66 to 17.67 GiB. Confirmed.

The narration says FP8 pushes the weights “into the high twenties”. A 27B model at one byte per parameter is 27 GB of weights. The published FP8 repository is 30.88 GB, which includes more than the weight tensors. The narration is describing the weights, so it is treated as sound, and no FP8 file size figure is put on screen.

Hardware guidance

Tier What is reported Source
16 GB GPU Possible with a ~10.7 GB two bit or ~13.4 GB three bit build, or by offloading part of a Q4 build to system RAM. Not a recommended configuration for this model. Qwen3.8-27B local hardware guide
12 GB laptop GPU with 32 GB system RAM, dynamic Q4 with CPU offload 3.26 tok/s on a long answer, 4.42 to 4.53 tok/s on short factual and coding answers as above
24 GB GPU Runs a good four bit build. Positioned as the practical everyday starting point. as above
32 GB GPU Breathing room. Q4 or Q5 with practical context. as above

RTX 3090, RTX 3090 Ti and RTX 4090 are all 24 GB cards, which is the class the video names at that tier.

Apple silicon

Machine Measured Source
M4 Mac Mini, 32 GB unified memory, MLX 4 bit 5 to 6 tokens per second, described as the only measured figure published for base M4 silicon Qwen3.8-27B on Apple silicon
M4 MacBook Air, 24 GB Fewer than ten tokens per second once a few thousand tokens in, memory bandwidth bound as above

The MLX four bit build is about 18 GB and needs roughly 16 to 19 GB of unified memory, which makes 24 GB the realistic floor on a Mac.

Generation speed on Apple silicon tracks memory bandwidth rather than capacity. Published bandwidth figures for the current line put M5 Ultra at 1.2 TB/s, M3 Ultra at 819 GB/s and M5 Max at 614 GB/s, all far above a base M4. This is the mechanism behind the video’s point that capacity lets a model load and bandwidth is what makes it feel alive.

Not found

No published measurement puts a 32 GB Mac at three to four tokens per second. The nearest measured figure for that machine is 5 to 6 tok/s on an M4 Mac Mini. The 3.26 to 4.53 tok/s range that does exist in the sources was measured on a 12 GB RTX 5070 Ti laptop using CPU offload, which is a different machine and a different failure mode. No token per second figure is put on screen anywhere in that chapter as a result.

Software named in the video

Name What it is
vLLM Open source inference and serving engine, OpenAI compatible endpoint, tensor parallelism, prefix caching, speculative decoding
SGLang Open source serving engine, listed by Qwen as supporting the YaRN extension to 1M context
GGUF The file format used by llama.cpp and the applications built on it
AWQ Activation aware weight quantization
YaRN Context window extension method
CUDA graphs NVIDIA mechanism for replaying a captured sequence of GPU work with less per launch overhead

Caveats