What Actually Fits: Named Models
Sorting llmfit's database by raw parameter count surfaces obscure or roleplay-oriented fine-tunes that happen to fit in memory, not the flagship models most people actually want to run. Querying llmfit info against each lab's own official repo (not a third-party quant mirror) gives a cleaner starting point, but llmfit's HuggingFace scrape has its own data-quality gaps (see the Kimi K3 row below, where it was off by roughly 2x). Every parameter count and MoE expert count in this table was cross-checked a second time against each lab's own model card, GitHub repo, or official announcement, not llmfit alone.
This table uses the current generation as of August 2026, not the mid-2024/2025 models (Llama 3.x, Qwen2.5, Mixtral, original DeepSeek-V3) that dated an earlier version of this page.
| Hardware memory budget | Model that fits | Architecture | Estimated weight residency at the stated precision or quantization |
|---|---|---|---|
| 16 GB VRAM (RTX 5060 Ti) or 16-32 GB unified (Mac mini M6) | openai/gpt-oss-20b (Aug 2025, Apache 2.0) fits with headroom; Qwen/Qwen3.8-27B only reaches this tier at a heavier quant (Q3_K_M), marginal at 99% memory utilization | MoE, 21B total, 3.6B active (4/32 experts) | 11.0 GB for gpt-oss-20b, 15.84 GB for Qwen3.8-27B at Q3_K_M |
| 32 GB VRAM (1x RTX 5090) or 32-48 GB unified (MacBook Pro M5 Pro) | Qwen/Qwen3.8-27B (Aug 2026, Apache 2.0) fits with large headroom; Qwen/Qwen3.6-35B-A3B (Apr 2026, Apache 2.0, MoE) also fits at 4-6-bit and decodes faster on unified-memory hardware because only ≈3B parameters activate per token | Dense ≈27B, or MoE 35B total / 3B active (256 experts, 8 routed + 1 shared per token) | 14.2 GB for Qwen3.8-27B; ≈20-26 GB for Qwen3.6-35B-A3B at 4-6-bit |
| 48 GB unified (MacBook Pro M5 Pro) | Qwen3.8-27B fits easily; meta-llama/Llama-4-Scout-17B-16E (109B total, MoE) does not quite fit | MoE, 17B active / 16 experts | 55.6 GB required, exceeds 48 GB |
| 64 GB VRAM (dual RTX 5090) | meta-llama/Llama-4-Scout-17B-16E | MoE, 109B total, 17B active | 55.6 GB for weights; it loads, but the remaining 8.4 GB is limited once runtime and KV cache are included |
| 96-128 GB (RTX PRO 6000, MacBook Pro M5 Max, DGX Spark, Ryzen AI Halo) | Llama-4-Scout fits with large headroom; no confirmed current-generation flagship lands specifically between 56 GB and 200 GB as of this snapshot | ||
| 192-256 GB (dual RTX PRO 6000, Mac Studio M5 Ultra) | deepseek-ai/DeepSeek-V4-Flash-0731 (MIT, GA release July 30-31, 2026) fits within the published third-party weight estimates on both configs; meta-llama/Llama-4-Maverick-17B-128E (≈400B total, MoE, community license) fits only on the 256 GB config | MoE, 304B total / 13B active per NVIDIA's official Build model card; the earlier preview build was 284B total at the same 13B active, and neither DeepSeek nor NVIDIA states why the GA release's total grew | No official VRAM figure exists. Third-party quantized estimates for the GA release cluster around 100-170 GB (Unsloth: ≈103 GB at 3-bit, ≈162 GB at "lossless" 8-bit; Spheron: ≈166 GB at INT4; llmfit: 155.8 GB). The upper end leaves about 22 GB before runtime and KV cache on a 192 GB system. Llama-4-Maverick-17B-128E needs 205.7 GB, so it fits 256 GB and does not fit 192 GB at that estimate |
| 192-256 GB (dual RTX PRO 6000, Mac Studio M5 Ultra, dual DGX Spark) | LibertAIDAI/GLM-5.3-Flash-NVFP4, a third-party NVFP4 quantization of Zhipu/Z.ai's GLM-5.3-Flash (MIT license, 1,048,576-token context, first natively multimodal model in the GLM-5 line: text, image, and video). Z.ai announced it officially on Aug 26, 2026, after running it anonymously as "Ox Alpha" on OpenRouter and OpenCode from Aug 20 | MoE, 320B total / 18B active across 45 layers, distinct from the 743B-parameter GLM-5.3 flagship SKU below | ≈181 GiB, down from 598.5 GiB in BF16. Weight-only NVFP4 on the routed FFN tensors (≈97% of parameters, the rest stays BF16); the checkpoint's authors report ≈0.99665 per-expert cosine round-trip fidelity against the BF16 source. Comfortably clears the 192 GB floor with headroom for runtime and KV cache, unlike the other rows in this bracket. The checkpoint was verified by its own authors only on a 2x DGX Spark pair (SGLang, tensor-parallel 2); their README states vLLM does not run on that combination |
| Any config on this page | zai-org/GLM-5.2 (≈753B total, MoE, MIT). GLM-5.3 (743B total, MoE, announced Aug 14, 2026, not Aug 17 as an earlier revision of this page stated) is the same base model with a post-training coding upgrade; weights ship staged, roughly two weeks after announcement. It is a separate SKU from GLM-5.3-Flash above, a smaller, distinct model that does fit this page's hardware once quantized | MoE, ≈40B active / 256 experts | 385.9 GB, exceeds everything here |
| Any config on this page | deepseek-ai/DeepSeek-V4-Pro-0813 (1.6T total, MoE, MIT, GA Aug 13, 2026) | MoE, 49B active (officially confirmed) | 845.4 GB, does not fit |
| Any config on this page | Qwen/Qwen3.8-2.4T-A95B (2.4T total, MoE, open-weight base of the hosted Qwen3.8-Max) | MoE, ≈95B active / 512 experts | 1,253 GB, does not fit |
| Any config on this page | MoonshotAI/Kimi-K3 (2.8T total, confirmed via Moonshot's own GitHub repo, Aug 2026) | MoE, 16 of 896 experts active (≈50B active, calculated) | ≈1,430 GB estimated (extrapolated from DeepSeek-V4-Pro's VRAM-per-parameter ratio; llmfit's own entry for this repo reports an incorrect 5,527B total and was not used) |
Sources for the officially-confirmed figures: gpt-oss-20b model card, Qwen3.8-27B, Qwen3.6-35B-A3B, Qwen3.8-2.4T-A95B, Llama 4 announcement, Llama-4-Scout model card, DeepSeek-V4-Flash-0731 model card and NVIDIA's Build model card for the 304B/13B figures (third-party VRAM estimates from Unsloth's deployment guide and Spheron's GPU recommender), DeepSeek-V4-Pro model card, GLM-5.2 announcement, Kimi K3 GitHub repo, GLM-5.3-Flash model card, Artificial Analysis GLM-5.3-Flash, GLM-5.3-Flash-NVFP4 checkpoint, Tutanka01's 2x DGX Spark deployment repo, kingjones30's 2x DGX Spark benchmark repo.
GLM-5.3-Flash-NVFP4's own maintainers publish no throughput number for the checkpoint, stating on their own repo that they "would rather publish nothing than publish a number we did not measure." Three third parties ran the checkpoint on the same class of hardware, a 2x DGX Spark pair, within two days of each other (Aug 26-27, 2026) and reported three different results. Tutanka01/glm5.3-flash-2x-dgx-spark-nvfp4, a single-author repo created Aug 26 with one star, uses SGLang and measured ≈18 tokens/sec end to end, with no benchmark method documented. kingjones30/GLM-5.3-Flash-2x-DGX-Spark, posted to an NVIDIA developer forum thread on Aug 27, uses a modified vLLM (NoPE-MLA zero-padding, Marlin MoE kernels) plus the checkpoint's own MTP-5 speculative decoding, and reports a three-run median of 24.74 tok/s for code, 30.30 tok/s for structured output, and 19.58 tok/s for prose; without speculative decoding, the same setup drops to a flat 14.6 tok/s, which its author attributes to the decode being bound by MoE memory bandwidth. A separate NVIDIA forum thread, titled with a "43.4 tok/s PEAK" claim, reports that single number with no benchmark method disclosed at all. The gap between 18, the 24.7-30.3 range, and 43.4 tok/s on the same two machines is most plausibly explained by engine choice (SGLang versus a hand-patched vLLM) and the presence or absence of MTP-5 speculative decoding, a hypothesis the two documented threads support but neither confirms outright. Treat all three numbers as community-reported field results, not independently verified benchmarks, and rerun the Benchmark Protocol section above before sizing a purchase around any of them.
The residency column above assumes every expert must be resident in VRAM or unified memory. It covers weights, not the complete runtime budget. llmfit reports both a "full model" figure and a much smaller "active" figure (for example DeepSeek-V4-Pro-0813 shows 845.4 GB full versus 63.6 GB active). The active figure describes the compute cost of a single forward pass; the full-model figure is what a naive all-in-accelerator-memory deployment needs. Runtime allocations, context length, KV-cache precision, batch size, and concurrent sequences consume additional memory. A shortcut does exist for MoE models, but it moves the constraint rather than removing it: the full expert set must stay resident somewhere fast (system RAM, not VRAM), and only the experts selected for the current token get streamed to the GPU or computed on CPU.
Two real projects implement exactly this. llama.cpp's --n-cpu-moe/--cpu-moe flags keep MoE expert tensors in CPU RAM while streaming the active ones to GPU over PCIe per token, confirmed via the project's own GitHub docs and issue discussions. FreeToken (Apache 2.0, arXiv paper, authors including Song Han and Ion Stoica) goes further with a bandwidth-adaptive hybrid mode: it profiles a machine's PCIe-vs-host bandwidth ratio (ft bench bw) and splits expert cache misses between "fetch over PCIe" and "compute in place on CPU," with an LRU cache of experts on GPU reusing whichever ones were active on the previous token. FreeToken's own README claims 3-4x faster decode and 6-30x faster prefill than Ollama on MoE models; that figure comes from the paper's authors, not an independent benchmark. FreeToken officially supports DeepSeek-V4-Flash, GLM-5.2, GLM-4.7, Qwen3.6/3.5 MoE, and gpt-oss, among others, and its ft launch claude command wires a locally-served model directly into Claude Code via FreeToken's Anthropic-compatible API. Community-reported (not independently verified) throughput numbers relayed on Slack: an 8 GB laptop GPU with 64 GB RAM running a 35B MoE model at 39.3 tokens/sec, an RTX 5090 with 192 GB RAM running DeepSeek-V4-Flash (284B) at 22 tokens/sec, and a 96 GB workstation GPU with 512 GB RAM running GLM-5.2 (753B) at 14.9 tokens/sec.
Practically, this means the "does not fit" verdicts for GLM-5.2 and DeepSeek-V4-Pro in the table above are true only for naive full-VRAM loading. With enough system RAM (512 GB-class, not the 96-256 GB VRAM figures this page uses elsewhere) and a CPU-offload-capable engine, both become usable at real, if reduced, throughput on hardware already covered in this page's fourteen configurations.
The frontier gap still widened rather than narrowed since the previous generation covered here: DeepSeek-V3 needed 350.6 GB at 4-bit, its August 2026 successor DeepSeek-V4-Pro needs 845.4 GB in full-VRAM terms, and neither Qwen's 2.4T-parameter MoE nor Moonshot's 2.8T-parameter Kimi-K3 fit on anything in this hardware lineup even with CPU offload, including the €30,000+ dual RTX PRO 6000 workstation. Reaching that class of model at usable speed requires either quantization aggressive enough that llmfit no longer rates it usable, or a budget and interconnect (NVLink-class, not the PCIe-only multi-GPU builds on this page) well past this page's scope.