BENCHMARKS × LOCAL MEMORY
Compare models on the Artificial Analysis benchmark that matters to you. Every local family stays on the graph: runnable models use their highest-fidelity comfortable build, while models beyond your Mac are muted at their smallest available build. The cloud rail traces Claude, Gemini, and GPT intelligence over time.
Qwen3.8 27B
Highest-scoring model, using its highest nominal-precision build that fits. Working estimate uses 45% of your 64 GB unified memory.
Artificial Analysis Intelligence Index v4.3: ten evaluations of reasoning, knowledge, coding, and agentic work. Refreshed September 18, 2026, with models released since read from the same index version on September 22; AA estimates are marked.
THE BENCHMARK MAP
COMFORTABLE FIT · PRIORITY
A dense vision-language Qwen model with 262K native context and MTP-trained decoding. Scores shown here come from Artificial Analysis.
THE FRONTIER GAP
FRONTIER CHANGELOG
Key changes to the Intelligence Index frontier, in release order.
MULTI-TOKEN PREDICTION
Qwen3.6 was trained with extra prediction heads. A compatible runtime can use them for speculative decoding: propose several future tokens, verify them together, and accept the ones the main model agrees with.
HOW GENERATION CHANGES
The MTP heads predict the next few tokens cheaply.
The full model checks those candidates in parallel.
Correct candidates advance generation several tokens.
This changes decoding, not the Artificial Analysis capability score. Rejected drafts are discarded, so Unsloth reports no accuracy change from enabling MTP.
faster token generation across Qwen3.6 MTP configurations
SAME MODEL, TWO DECODING PATHS
♥ Loading Hugging Face likes…
Both repositories contain Qwen3.6 35B-A3B. They share the same capability score; the MTP build adds prediction heads that a compatible runtime can use to accelerate generation.
+0.49 GBextra on disk for UD-Q4_K_XLBaseline decoding. No MTP-specific single-slot or projector restriction.
♥ Loading likes…
hf download unsloth/Qwen3.6-35B-A3B-GGUF --include "*UD-Q4_K_XL*.gguf" --local-dir "./models/Qwen3.6-35B-A3B-GGUF"MTP-aware speculative decoding with two draft tokens. Fit adds approximately 1 GB of runtime headroom.
♥ Loading likes…
hf download unsloth/Qwen3.6-35B-A3B-MTP-GGUF --include "*UD-Q4_K_XL*.gguf" --local-dir "./models/Qwen3.6-35B-A3B-MTP-GGUF"File sizes are summed from matching GGUFs in both Hugging Face repositories, checked July 26, 2026. The disk delta is exact for the selected artifact; speed and extra runtime memory are Unsloth-reported ranges. Like counts refresh from the Hugging Face Hub API and are summed across repositories, not deduplicated users.
Speed is hardware-dependent. Unsloth recommends trying draft counts 1–6, although two performed best on average. Their measured acceptance fell from 83% at two drafts to 50% at four, making extra drafts less useful.
Current llama.cpp support requires MTP-aware settings. Parallel slots above one and the multimodal projector are not yet supported with this path.
DOWNLOADABLE BUILDS
Actual published artifact sizes from unsloth/Qwen3.8-27B-GGUF. Fit reserves 20% of unified memory for macOS and inference overhead.
Q4_K_M
Q6_K
Q8_0
BF16
File size is not peak runtime memory. Context length, KV cache, multimodal projectors, and the inference engine add overhead. The benchmark score belongs to the base model—not to each quantized file.
THE FULL MTP REPOSITORY
Filter by nominal bit depth. “Working estimate” adds Unsloth’s recommended ~1 GB MTP headroom to the actual file size.
DOWNLOAD UD-IQ2_XXS
Requires the Hugging Face CLI. The wildcard fetches every shard when a GGUF is split across files.
hf download unsloth/Qwen3.6-27B-MTP-GGUF --include "*UD-IQ2_XXS*.gguf" --local-dir "./models/Qwen3.6-27B-MTP-GGUF"Unsloth Dynamic: model-specific mixed precision, assigning more bits to sensitive layers.
Importance-aware quantization, designed to retain more useful signal at very low bit counts.
llama.cpp superblock quantization. S, M, and XL distinguish recipes that spend progressively more storage.
Older uniform 4-bit formats. Widely compatible, but less selectively allocated than newer recipes.
Do not rank these files by tokens/second from size alone. Unsloth warns its per-quant throughput measurements are noisy; MTP draft acceptance, hardware bandwidth, and runtime settings can matter more than a small size difference.
CURRENT VIEW
Open models with both a Intelligence Index score and a published Q4_K_M artifact, sorted by the selected benchmark.
READ THE FINE PRINT
The selector changes the independent model evaluation shown on the vertical axis. Missing evaluations are omitted, never treated as zero. Intelligence scores use the v4.3 index, refreshed September 18, 2026, with models released since added from the same index version on September 22. Values estimated by Artificial Analysis are marked “AA estimate”; inherited quantized-model scores are marked “base score.” Coding and Agentic are archived September 1 scores, and remain separate from the current index.
Artifact sizes are the summed byte sizes of the named files in the linked repositories, including multi-part builds and native quantized checkpoints.
Artificial Analysis evaluates a model endpoint or base model. The app does not pretend every Q4, Q6, or Q8 artifact preserves that score exactly.
“Comfortable” keeps 20% of unified memory free. Actual runtime needs still vary with context and cache precision. Each family appears once: the highest-fidelity comfortable artifact when one fits, otherwise its smallest available artifact in gray.
Benchmark data © Artificial Analysis and used with attribution. Snapshot captured July 2026. Artificial Analysis’ API requires a server-side key; full evaluation fields depend on API tier.