OrcaSAQ-2 is a small EXL3 quantization of Qwen3.8-27B with publisher-reported WikiText-2 perplexity close to the BF16 reference. That does not prove it matches BF16 on other tasks or outperforms other quantizations: the available alternatives were evaluated under different conditions. Compare the files, runtime, vision support and results from tests that match your own workload.
What OrcaSAQ-2 changes
OrcaRouter’s model card describes OrcaSAQ-2-27B as an EXL3 quantization of Qwen3.8-27B, with an average 3.21 bits per decoder weight. The listed checkpoint is 12.3 GB, compared with 54 GB for the BF16 reference. It supports thinking mode, tool calling and MTP speculative decoding.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
GMKtec EVO-X2 AI Mini PC Ryzen Al Max+ 395 Superchip 128GB LPDDR5X 2TB SSD | $3,649.99 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
The official Qwen3.8 project repository points to the Qwen weights and model card. Qwen describes Qwen3.8-27B as a dense 27B model with image and video understanding, 64 layers and a native context length of 262,144 tokens, extensible to one million tokens. OrcaSAQ-2 does not include the base model’s visual encoder, so the quantized checkpoint itself is text-only.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →What the published quality results show
OrcaRouter reports a comparison between OrcaSAQ-2 and the BF16 reference using the same evaluation path and 16,376 predicted tokens from WikiText-2. The reported perplexity is 5.6482 for OrcaSAQ-2 and 5.6468 for BF16, a +0.02% change according to the publisher. OrcaRouter also reports a mean KLD of 0.031 and 93.2% top-1 token agreement with the reference. These are publisher-reported figures from that test, not independent results.
#1 Best Overall
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Perplexity summarizes how well a model predicts text across a dataset; top-1 agreement measures how often its most likely next token matches the reference. Similar aggregate perplexity can therefore coexist with different token choices. The WikiText-2 result is evidence about language modeling on that evaluation, not proof that downstream answers, reasoning, coding or agent behavior are unchanged. Local Model Watch’s analysis likewise notes that the available card does not provide same-condition tests against GGUF or other EXL3 quantizations.
How it compares with other published quantizations
ISTA-DASLab publishes Qwen3.8-27B GSQ-RCO in GGUF at 2.50, 2.75, 3.00 and 3.50 bpw, with listed file sizes from 8.4 to 11.8 GB. Its card reports evaluations against BF16 and Unsloth Dynamic versions of the base model. The following figures are useful as separate evidence, but they are not a head-to-head comparison with OrcaSAQ-2:
| Option and source | Format and listed footprint | Reported evaluation |
|---|---|---|
| OrcaSAQ-2-27B, OrcaRouter model card | EXL3; 3.21 average decoder bpw; 12.3 GB checkpoint | Against BF16 on WikiText-2, 16,376 predicted tokens: perplexity 5.6482 versus 5.6468; 93.2% top-1 agreement and mean KLD 0.031. |
| GSQ-RCO IQ3_S, ISTA-DASLab model card | GGUF; 3.50 bpw; 11.8 GB | In that card’s evaluation, AIME25 100.00 and LiveCodeBench v6 85.71, matching its BF16 figures; GPQA-Diamond 89.39 versus BF16’s 89.90. |
Sources: OrcaRouter and ISTA-DASLab. The table’s scores come from different test suites and protocols; they should not be read as a ranking between the two quantizations. The GSQ-RCO card’s results describe its own test setup, while the OrcaSAQ-2 figures above are a WikiText-2 comparison with BF16.
Recommended Free Tools
Why the other available runs do not settle the ranking
A separate community comparison of FP8 and INT4/INT8 AutoRound checkpoints reports results under a shared workload but describes the comparison as only partially comparable. The model checkpoint and quantization vary together, quality results are single trials, and the FP8 throughput run used a tokenizer fallback. It is exploratory workload-specific evidence, not an isolated estimate of what quantization alone changes.
No cited source supplies a controlled, repeated comparison of OrcaSAQ-2 against the alternatives using the same hardware, harness and settings. OrcaRouter itself cautions that public agent scores use different agent stacks and should not be interpreted as a strict model-only ranking. Avoid combining unrelated benchmark scores into a league table.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Which Qwen3.8-27B quantization fits your GPU?
Check more than the checkpoint file
The listed file size is not the full memory budget. Runtime overhead, KV cache, batch size and requested context length also consume memory, so a smaller checkpoint does not guarantee that a particular serving setup will fit. Qwen’s 262,144-token native context is an architecture capability, not a promise that a GPU can serve that length with a chosen cache and batch configuration.
OrcaRouter says its measurements were made under a 15.7 GiB GPU memory cap and describes around 32K interactive context as a practical starting point on a 16 GB GPU. Treat that as the publisher’s guidance, not a universal fit guarantee: confirm memory use with your runtime, context and serving pattern.
Match the format to your deployment stack
OrcaSAQ-2 is EXL3, and OrcaRouter provides vLLM instructions. The cited GSQ-RCO alternatives are GGUF and list use with llama.cpp, Ollama and LM Studio. If you already operate one of these stacks, compatibility and setup may matter more than a small difference in a score from a different evaluation.
Decide whether you need image or video input
OrcaSAQ-2 omits the visual encoder that gives the base Qwen3.8-27B its image and video capability. The GSQ-RCO card lists a separate BF16 vision projector for multimodal use. If vision matters, verify that the companion artifact and your chosen runtime are supported; do not assume a text quantization alone retains the base model’s visual input path.
Compare task quality on a matched workload
For a useful local test, compare the same underlying Qwen3.8-27B tasks with the same prompts, harness, decoding settings, hardware and runtime. Repeat trials where possible, especially for tasks with variable outputs. Keep perplexity, token agreement, reasoning benchmarks, coding tests and long-horizon agent evaluations separate: each measures a different aspect of behavior.
What the serving figures can—and cannot—tell you
OrcaRouter reports 65.3 tokens per second at one stream without MTP and 90.1 tokens per second with MTP enabled, under its stated 15.7 GiB memory cap. These are publisher measurements, not independent benchmark results. The same card reports that aggregate throughput at eight and 16 streams is lower with MTP enabled; it says MTP consumes KV capacity and advises benchmarking both settings for highly batched workloads.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Those figures suggest testing MTP against your own traffic pattern rather than assuming it will improve throughput at every concurrency level. Results from a single stream do not establish performance for a busy multi-stream server.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
