{"slug": "model-metis-1-6-praxis", "title": "Model: Metis-1.6 Praxis", "summary": "Lernex released Metis-1.6 Praxis, the smaller of its two Metis-1.6 models, on 1 November under Apache 2.0, and llama.cpp has run it since 3 November. Metis-1.6 Praxis is a 2.95B-parameter looped hybrid of 12 layers — mostly Mamba-2 with two attention layers and a small mixture-of-experts block on every layer — that runs its stack twice per token, so each token reads about 1.1B parameters' worth of weights and keeps about 8 KB of attention cache per token across a 128K window. Lernex's GGUF files range from 1.84 GB (IQ4_XS) to 3.22 GB (Q8_0), with Q4_K_M needing about 2.2 GB of VRAM at 4K context and 3.2 GB at 128K, and no per-level quality loss has been published.", "body_md": "# Metis-1.6 Praxis\n\n3B MoE\nMetis\n\nThe smaller of Lernex's two Metis-1.6 models, released on 1 November under Apache 2.0. It is a looped hybrid: 12 layers, mostly Mamba-2 with two attention layers and a small mixture-of-experts block on every layer, and the whole stack runs twice per token, so each token reads about 1.1B parameters' worth of weights from a 2.95B model. Each pass keeps its own attention cache, about 8 KB per token in total, which makes its 128K window cheap. At Q4_K_M it needs about 2.2 GB at 4K and 3.2 GB at 128K, comfortable on a 4 GB card. Q8_0 needs about 4.5 GB at 128K, comfortable on a 6 GB card. llama.cpp runs it since 3 November. Sizes come from Lernex's own GGUF files, and no per-level quality loss has been published.\n\n131K\n\nMax Context\n\n5\n\nQuant Variants\n\nGGUF Q8_0\n\nBest Quality\n\n—\n\nPerplexity rise (best build)\n\n## Quantization Variants\n\nPer-quant VRAM, quality loss, and inference speed on RTX 4090\n\nMeasured = run on this site · Estimated = not run here: calculated, or a figure this site has not verified · Community = public reports\n\n| Format | Level | BPW | VRAM | PPL Loss | Speed | Source | Actions | \n|---|---|---|---|---|---|---|---|\n| GGUF | IQ4_XS | 4.98 | 1.9 GB | — | — | Estimated |  | \n| GGUF | Q4_K_M | 5.48 | 2.1 GB | — | — | Estimated |  | \n| GGUF | Q5_K_M | 6.31 | 2.5 GB | — | — | Estimated |  | \n| GGUF | Q6_K | 7.04 | 2.8 GB | — | — | Estimated |  | \n| GGUF | Q8_0 | 8.74 | 3.4 GB | — | — | Estimated |  | \n\n### Which file each size is for\n\nA file size is not a memory figure: the VRAM column is the calculator's total at 4K context — weights plus KV cache plus a 10% activation buffer. Units differ on purpose: file sizes are decimal GB, as Hugging Face lists them; the VRAM column is GiB (1024³ bytes) with 2% added to the weights, so 7.12 GB of file counts as about 6.8 GiB there. A vision projector (mmproj), where the repo ships one, is a separate download and is in neither number.\n\n- GGUF IQ4_XS[Lernex/Metis-1.6-Praxis-GGUF](https://huggingface.co/Lernex/Metis-1.6-Praxis-GGUF) · Metis-1.6-Praxis-IQ4_XS.gguf · 1.84 GB (1,836,154,464 bytes) · listing read 2026-11-04\n- GGUF Q4_K_M[Lernex/Metis-1.6-Praxis-GGUF](https://huggingface.co/Lernex/Metis-1.6-Praxis-GGUF) · Metis-1.6-Praxis-Q4_K_M.gguf · 2.02 GB (2,019,139,168 bytes) · listing read 2026-11-04\n- GGUF Q5_K_M[Lernex/Metis-1.6-Praxis-GGUF](https://huggingface.co/Lernex/Metis-1.6-Praxis-GGUF) · Metis-1.6-Praxis-Q5_K_M.gguf · 2.33 GB (2,325,706,336 bytes) · listing read 2026-11-04\n- GGUF Q6_K[Lernex/Metis-1.6-Praxis-GGUF](https://huggingface.co/Lernex/Metis-1.6-Praxis-GGUF) · Metis-1.6-Praxis-Q6_K.gguf · 2.60 GB (2,597,131,872 bytes) · listing read 2026-11-04\n- GGUF Q8_0[Lernex/Metis-1.6-Praxis-GGUF](https://huggingface.co/Lernex/Metis-1.6-Praxis-GGUF) · Metis-1.6-Praxis-Q8_0.gguf · 3.22 GB (3,224,393,312 bytes) · listing read 2026-11-04\n\n## Similar models\n\n[Compare with Metis-1.6 Logos](https://quantized.uk/tools/compare/?a=metis-1-6-praxis&b=metis-1-6-logos)\n\n[10B MoEConsumer GPUMac / Apple SiliconQ4_K_M7.1 GB—perplexity rise](https://quantized.uk/quant-hub/metis-1-6-logos/)\n\n### Metis-1.6 Logos\n\nMetis\n\nThe larger of Lernex's two Metis-1.6 models, released on 1 November under Apache 2.0. It is a looped hybrid: 20 layers, mostly Mamba-2 with three attention layers and a mixture-of-experts block on every layer, and the whole stack runs twice per token, so each token reads about 2.6B parameters' worth of weights from a 10.24B model. Each pass keeps its own attention cache, about 15 KB per token in total. IQ4_XS needs about 6.4 GB at 4K and 6.9 GB at 32K, comfortable on an 8 GB card. Q4_K_M needs about 7.1 GB at 4K, tight on an 8 GB card and comfortable from 10 GB up; at 128K it needs about 9.1 GB, comfortable on a 12 GB card. Q8_0 needs about 11.7 GB at 4K, comfortable on a 16 GB card. llama.cpp runs it since 3 November. Sizes come from Lernex's own GGUF files, and no per-level quality loss has been published.\n\n[3BConsumer GPUMac / Apple SiliconQ4_K_M2.5 GB+3.5%perplexity rise](https://quantized.uk/quant-hub/llama-3.2-3b/)\n\n### Llama 3.2 3B Instruct\n\nMeta Llama 3.2\n\nTiny but capable. Runs on 4GB VRAM or 8GB RAM, even on phones via llama.cpp.\n\n[3BConsumer GPUMac / Apple SiliconQ4_K_M2.1 GB+3.2%perplexity rise](https://quantized.uk/quant-hub/qwen2.5-3b/)\n\n### Qwen2.5 3B Instruct\n\nAlibaba Qwen2.5\n\nTiny Qwen2.5 for edge devices. Runs on 4GB VRAM or Raspberry Pi class hardware.\n\n[1BConsumer GPUMac / Apple SiliconQ4_K_M0.9 GB+4.2%perplexity rise](https://quantized.uk/quant-hub/llama-3.2-1b/)\n\n### Llama 3.2 1B Instruct\n\nMeta Llama 3.2\n\nUltra-light Llama for mobile and embedded. Sub-2GB VRAM with Q4.\n\n## Running Metis-1.6 Praxis locally\n\nAt Q4_K_M and 4K of context, Metis-1.6 Praxis needs about 2.1 GB — 1.9 GB of weights, 0.03 GB of KV cache and a 0.2 GB activation buffer. The smallest cards in this index that clear that comfortably are 8 GB ones — 11 of them, from the Mac M1 8G up — and 78 of the 78 cards here do. These are calculated figures, not measurements: the estimate stops counting a card as comfortable at 88% of its VRAM, which is roughly the room a desktop session needs.\n\n## What longer context costs\n\nGoing from 4K to 32K adds about 0.22 GB, taking the total to 2.4 GB. That is gentler than the parameter count suggests: only 4 of 12 layers keep a KV cache that grows with context, so the cache scales at a fraction of the usual rate. The model's native window is 128K; holding all of it at Q4_K_M would need about 3 GB.\n\n## Which build to download\n\nThis index only tracks Metis-1.6 Praxis in GGUF, across 5 levels (IQ4_XS, Q4_K_M, Q5_K_M, Q6_K, Q8_0). No per-level perplexity sweep has been published for this model, so the quality column is empty rather than estimated — within one model, more bits per weight is the only ordering the data supports. No throughput has been measured for this model on this site's hardware. GGUF runs on llama.cpp and Ollama across NVIDIA, AMD and Apple silicon; AWQ and GPTQ target vLLM on CUDA and ROCm; EXL2 is ExLlamaV2 — CUDA only, and now archived, so local use rather than serving.\n\n## Common questions\n\n- How much VRAM does Metis-1.6 Praxis need?\n- About 2.1 GB at Q4_K_M with 4K of context and batch 1, rising to roughly 2.4 GB at 32K. That figure is weights plus KV cache plus a 10% activation buffer, calculated from the model's architecture rather than measured on a card.\n- Will Metis-1.6 Praxis run on a 8GB GPU?\n- Yes — at Q4_K_M and 4K context it needs about 2.1 GB, which leaves 3.9 GB spare on a Mac M1 8G, or on any of the other 8 GB cards here (11 in all). That is the smallest size in this index that clears it comfortably; 78 of 78 do.\n- Which quantization of Metis-1.6 Praxis should I use?\n- No published quality comparison exists for this model, so pick by footprint: Q4_K_M is the level most people run, and a higher bits-per-weight level is more faithful. All 5 levels in the index are IQ4_XS, Q4_K_M, Q5_K_M, Q6_K, Q8_0. Note this is a mixture-of-experts model — all parameters must be resident even though only a fraction are active per token, so the memory cost follows the total, not the active count.\n\n## Where this model fits\n\nSized at Q4_K_M with a 4K context window, smallest card first. Comfortable means the estimate uses at most 88% of the memory.\n\n- Runs comfortably on\n- [Mac M1 8G8GB · +3.9](https://quantized.uk/gpu/mac-m1-8g/)[Mac M3 8G8GB · +3.9](https://quantized.uk/gpu/mac-m3-8g/)[RTX 40608GB · +5.9](https://quantized.uk/gpu/rtx-4060/)[RTX 4060 Ti 8G8GB · +5.9](https://quantized.uk/gpu/rtx-4060-ti-8g/)[Radeon RX 9060 XT 8G8GB · +5.9](https://quantized.uk/gpu/radeon-rx-9060-xt-8g/)\n- Ships in\n- [GGUF](https://quantized.uk/formats/gguf/)\n- Try it yourself\n- [Size it in the calculator](https://quantized.uk/tools/vram-calc/?model=metis-1-6-praxis&quant=Q4_K_M)[Generate the run command](https://quantized.uk/tools/cli-gen/?model=metis-1-6-praxis&quant=Q4_K_M)\n\n## How to actually run this\n\nDeployment guides for this model and this class of hardware.\n\nThis page's figures change when the model or the runtime does.Last updated 2026-11-05 [RSS → /feed.xml](https://quantized.uk/feed.xml)", "url": "https://wpnews.pro/news/model-metis-1-6-praxis", "canonical_source": "https://quantized.uk/quant-hub/metis-1-6-praxis/", "published_at": "2026-11-05 12:00:00+00:00", "updated_at": "2026-11-05 12:23:56.257510+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-tools", "machine-learning"], "entities": ["Lernex", "Metis-1.6 Praxis", "Metis-1.6 Logos", "llama.cpp", "Hugging Face", "Mamba-2"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/model-metis-1-6-praxis", "markdown": "https://wpnews.pro/news/model-metis-1-6-praxis.md", "text": "https://wpnews.pro/news/model-metis-1-6-praxis.txt", "jsonld": "https://wpnews.pro/news/model-metis-1-6-praxis.jsonld"}}