BEST OPEN-WEIGHT LLM
BY MEMORY REQUIRED

A field guide to freely downloadable models across ten hardware classes — Raspberry Pi to a half-terabyte Graviton4 box. Quantized GGUF (llama.cpp) as the common denominator; ONNX / vLLM noted where relevant.

compiled 2026-07-24 · grounded in ~40 web sources (§6) · • measured / community-reported · ≈ bandwidth-scaled estimate · prices = US street / AWS on-demand

TL;DR
  • Licenses stopped being a problem. Nearly everything worth running is Apache-2.0 (Qwen 3.5/3.6, Gemma 4, gpt-oss) or MIT (DeepSeek V4, GLM-5.2, Phi-4-mini). Gemma 4 moved to Apache-2.0 in Apr '26; the Llama Community license (700M-MAU cap) is now the odd one out.
  • Sparse MoE ate the local stack. You pay memory for total parameters but bandwidth only for active ones (3–13B). Qwen3.6-35B-A3B, Gemma 4 26B-A4B, gpt-oss-20b/120b and DeepSeek V4-Flash are all this shape — it is why a $200 RTX 3060 or a plain i9 now runs 30B-class models at chat speed.
  • Defaults by tier: Pi → Llama 3.2 3B / Gemma 4 E2B · 16 GB CPU → gpt-oss-20b · 32 GB CPU → Qwen3.6-35B-A3B · 12–24 GB GPU → Qwen3.6-27B or 35B-A3B · 32 GB (5090) → dense 32B + gpt-oss-120b via expert offload · big-RAM Graviton4 → gpt-oss-120b or DeepSeek V4-Flash.
  • Decode speed is a bandwidth quotient, not a core count: t/s ≈ effective GB/s ÷ bytes of active weights read per token (× 0.2–0.6 real-world efficiency). The ladder below is most of the story.
  • Hardware, not models, is the 2026 bottleneck. The AI-driven DRAM shortage has a Pi 5 16 GB at ~$205–300 (was $120) and RTX 5090 street price at $3.7–4.3K vs a $2.0K MSRP.
SCOPE & METHOD

Included: models with weights freely downloadable (Hugging Face) and runnable on open-source inference stacks — llama.cpp/GGUF first, plus ONNX Runtime GenAI, Ollama, vLLM. Best 2–3 per class chosen on benchmark quality per GB, community adoption, and measured or estimable throughput on the class hardware. Q4_K_M-class quantization assumed unless noted (MXFP4 for gpt-oss). All t/s figures are single-stream decode; treat as planning numbers — build, context fill and RAM tuning swing results ±2×.

MARKET CONTEXT — THE 2026 DRAM CRISIS

AI datacenter buildout absorbed the memory supply: LPDDR4 up ~7× in a year (Raspberry Pi), contract DRAM forecast +up to 95% in Q1'26 (TrendForce), NVIDIA cut RTX-50 output 20–40% and let the 5090 drift from $1,999 MSRP to $3.7–4.3K street. Pi 5 16 GB went $120 → $205 (Feb) with further rises in April. Consequence: picking a model is now mostly a memory-price question — which is exactly how this report is organized.

WHAT CHANGED, FEB–JUL 2026

Qwen 3.5 (Feb, Apache 2.0, 0.8B→397B) and Qwen 3.6 (Apr: 27B dense + 35B-A3B) · Gemma 4 (Apr 2, first Gemma under Apache 2.0; E2B→31B) · DeepSeek V4 (Apr 24, MIT: Flash 284B-A13B, Pro 1.6T-A49B) · GLM-5.2 (Jun, MIT — top open GPQA) · Kimi K2.7 Code (Jun, Modified-MIT, ~1T) · MiniMax M3 · Kimi K3 (2.8T) weights due Jul 27. Frontier-adjacent quality now ships MIT/Apache, and everything ≤ ~13B active runs on consumer gear.

THE ONLY NUMBER THAT MATTERS — memory bandwidth (GB/s, log scale). Decode t/s ≈ bandwidth ÷ bytes of active weights per token.
Memory bandwidth ladder (GB/s, log scale)Raspberry Pi 517LPDDR4XDDR5-6000 2-ch96desktop CPUStrix-Halo-class256LPDDR5XArc B580456GDDR6Graviton4 /socket53712-ch DDR5RX 9070 XT640GDDR6RTX 3090936GDDR6XRTX 40901008GDDR6XRTX 50901792GDDR7

2TEN HARDWARE CLASSES — THE PICKS

Each card: hardware definition and Jul-'26 price, the 2–3 models to run, memory footprint at ~4-bit, expected single-stream decode speed on the class hardware (• measured, ≈ estimated), and why. Required anchor classes — Raspberry Pi 4, i9 at 16 GB, RTX 30/40/50 — are C1, C3, C5, C6, C7.

C1EDGE ARM SBCRaspberry Pi 4 / Pi 5 · 4× Cortex-A72/A76 · 8–16 GB LPDDR4X · ~17 GB/sPi 5 8 GB ≈$125 · 16 GB ≈$205–300 · Pi 4 8 GB ≈$85–115
ModelFootprint (~Q4)Est. decode t/sWhy this pick
Llama 3.2 3B Instruct~2.0 GB• 8–9 (Pi 5) · ~3–4 (Pi 4)Best-tested small model on Pi; reliable structure/tool output; reading-speed generation on a Pi 5.
Gemma 4 E2B~2.5–3 GB≈ 6–10 (Pi 5)The only multimodal option in class: image + native audio input, 128K ctx, Apache-2.0.
Qwen3.5-4B~2.5 GB≈ 4–7 (Pi 5)Highest quality that stays usable; thinking-mode toggle; 119-language coverage.
NB 7–8B dense fits in 8 GB but decodes at ~1–3 t/s — below reading speed. Pi 4 runs at ~40–50% of Pi 5. 1B-class (Llama 3.2 1B, TinyLlama) hits • 12–18 t/s when you only need routing/classification.
C2MAINSTREAM x86 CPUIntel Core i5/i7 (12400 → 14700), iGPU only · 8–16 GB DDR4/DDR5 · ~50–95 GB/scomplete systems ≈$500–900 — the class most people already own
ModelFootprint (~Q4)Est. decode t/sWhy this pick
gpt-oss-20b (MXFP4)12.9 GB file · 14–16 GB RAM≈ 10–20The 16-GB headline act: o3-mini-class reasoning + native tool use on a stock desktop, no dGPU.
Qwen3.5-9B~5 GB≈ 8–14Default dense all-rounder; leaves RAM for the rest of the machine.
Gemma 4 12B Unified~7 GB≈ 6–102026's laptop model: text+image+video+audio input under Apache-2.0.
NB At 8 GB RAM drop to the 4B class (Qwen3.5-4B, Gemma 4 E4B, Phi-4-mini 3.8B/MIT) — same picks as C4, minus the speed.
C3HIGH-END DESKTOP CPUIntel Core i9-13900K/14900K (24c) · 16–32 GB DDR5 · ~90 GB/s dual-channelCPU ≈$450–550 · full system ≈$1.3–1.8K
ModelFootprint (~Q4)Est. decode t/sWhy this pick
Qwen3.6-35B-A3B~21 GB · needs 32 GB• 12–25 (30+ tuned)CPU MoE king: GPQA-Diamond 86.0 while reading only ~3B params per token.
gpt-oss-20b (MXFP4)12.9 GB · fits 16 GB≈ 15–25The pick at exactly 16 GB; adjustable reasoning effort, agent-grade function calling.
Gemma 4 26B (A4B MoE)≈16 GB≈ 12–20Apache MoE with vision input; top-10 open model on Arena (text).
NB i9 @ 16 GB (anchor point): gpt-oss-20b is the ceiling; 32 GB unlocks all three. Dense 27–32B decodes at ~3–5 t/s on CPU — skip it. Bandwidth, not cores, is the wall: XMP left off has cost users 3× throughput.
C4ENTRY NVIDIAGTX 1650 · RTX 3050 · RTX 4060 · 4–8 GB VRAM · 192–272 GB/sused 4 GB cards <$150 · RTX 4060 8 GB ≈$300+ ('26-volatile)
ModelFootprint (~Q4)Est. decode t/sWhy this pick
Qwen3.5-4B~2.5 GB≈ 60–902024-flagship quality in toy VRAM; room for 16K+ ctx even on 6 GB.
Qwen3.5-9B / Llama 3.1 8B~5 GB (8 GB cards)• 40–55The workhorse tier: chat, RAG, light coding — fully resident on an 8 GB card.
Gemma 4 E4B~3 GB≈ 45–70Vision + audio input in ~3 GB; Apache-2.0.
NB On 4 GB stay ≤4B — or run Qwen3.6-35B-A3B with experts offloaded to system RAM (llama.cpp --n-cpu-moe): slower, but a 35B on a $150 card.
C5NVIDIA RTX 30-SERIES3060 12 GB → 3080 → 3090 24 GB · GDDR6/6X · 360–936 GB/s3060 12 GB ≈$200 used · 3090 24 GB: used market, '26-volatile
ModelFootprint (~Q4)Est. decode t/sWhy this pick
Qwen3.6-35B-A3B21–22 GB (3090)• 101 (3090) · • 38 (3060 + expert offload)The best perf-per-dollar trick of 2026: a $200 Ampere runs a 35B at chat speed.
Qwen3.6-27B~16–17 GB≈ 30–35 (3090)SWE-bench Verified 77.2 — best open dense coder, fully resident in 24 GB.
gpt-oss-20b (MXFP4)12.9 GB · wants ≥15 GB VRAM≈ 150–180 (3090) · • 30 (3060, spills)Reasoning at interactive speed on 3080 Ti / 3090-class cards.
NB 12 GB sweet spot: Qwen3 14B Q4 (~9 GB, ≈25–30 t/s) or Gemma 4 12B for multimodal. gpt-oss-20b under ~15 GB VRAM spills experts to RAM and drops to ~30 t/s.
C6NVIDIA RTX 40-SERIES4070 Super 12 GB → 4070 Ti Super 16 GB → 4090 24 GB · 504–1008 GB/s4070 Ti Super ≈$799 · 4090 (EOL) refurb ≈$2.5–2.8K
ModelFootprint (~Q4)Est. decode t/sWhy this pick
gpt-oss-20b (MXFP4)12.9 GB• 225 (4090)Agentic loops in real time; throughput champion of the 24 GB tier.
Qwen3.6-27B~16–17 GB≈ 38–45 (4090)Daily-driver coder; 256K ctx stays cheap (hybrid linear attention → small KV).
Gemma 4 31B≈19 GB≈ 33–40 (4090)MMLU-Pro 85.2 · AIME'26 89.2 — strongest open generalist that fits in 24 GB.
NB 12 GB 40-cards behave like C5's 3060 tier, ~1.5–2× faster (4070 Super: ~75 t/s on 8B). At 16 GB (4060 Ti 16 / 4070 TiS) gpt-oss-20b runs fully resident.
C7NVIDIA RTX 50-SERIES5070 Ti / 5080 16 GB → 5090 32 GB · GDDR7 · 896–1792 GB/s5070 Ti ≈$775–970 · 5080 ≈$1.1–1.35K · 5090 ≈$3.7–4.3K street (MSRP $2.0K)
ModelFootprint (~Q4)Est. decode t/sWhy this pick
Qwen3 32B (dense)~19–20 GB• 61 (5090)The 32-GB payoff: a dense 32B fully in VRAM with ~32K ctx at interactive speed.
gpt-oss-120b (MXFP4)60.8 GB · experts → DDR5≈ 25–45 (5090 + fast RAM)o4-mini-class at home; full 128K ctx workable via MoE expert offload.
Qwen3.6-35B-A3B21–22 GB≈ 120–190 (5090)Plus • 282 t/s on gpt-oss-20b — the fastest single-card local stack, period.
NB 16 GB 50-cards (5070 Ti / 5080): gpt-oss-20b resident; 35B-A3B with offload ≈30–50 t/s. A 5090 cannot hold a dense 70B at Q4 (~40 GB) — drop to Q3 or accept offload.
C8AMD RADEONRX 9070 XT 16 GB (RDNA4) · RX 7900 XTX 24 GB (RDNA3) · 640 / 960 GB/s9070 XT ≈$649 · 7900 XTX ≈$800–1.1K
ModelFootprint (~Q4)Est. decode t/sWhy this pick
gpt-oss-20b (MXFP4)12.9 GB• 102 (XTX)CUDA-free reasoning at ~36% of a 5090's speed for ~22% of its street price.
Qwen3.5-9B / Llama 3.1 8B~5 GB• 96 (XTX)Daily driver on ROCm 7.2 or the zero-setup Vulkan backend.
Qwen3.6-35B-A3B21–22 GB≈ 75–90 (XTX)24 GB holds the 2026 default MoE fully resident.
NB ROCm 7.2 (Mar '26): first release with out-of-box Ollama / LM Studio / llama.cpp / vLLM parity incl. RDNA4 — still ~10–25% behind CUDA on equal silicon. 9070 XT ≈ 16-GB picks at ~2/3 XTX speed.
C9INTEL ARCB580 12 GB (Battlemage) · A770 16 GB (Alchemist) · 456 / 560 GB/sB580 $249 MSRP · A770 16 GB ≈$279 — cheapest VRAM per dollar on the market
ModelFootprint (~Q4)Est. decode t/sWhy this pick
Qwen3.5-9B / Llama 3.1 8B~5 GB• 35–55 (A770, SYCL) · ≈ 50–80 (B580)The budget 8B box; B580 numbers vary strongly by stack.
Qwen3 14B / Qwen2.5-14B~9 GB• 21 (A770)Meaningfully smarter than 8B, with ~6 GB headroom for 8K ctx on the A770.
gpt-oss-20b (MXFP4)12.9 GB → A770 16 GB≈ 25–40The reasoning tier on a $279 card.
NB IPEX-LLM was archived Jan '26 — current paths: llama.cpp SYCL or Vulkan, Ollama Vulkan, OpenVINO. Expect more stack variance than on NVIDIA/AMD.
C10AWS GRAVITON4 · CPU, LARGE-MEMr8g / x8g · Neoverse-V2, up to 96c/socket · 192 GB–3 TB · 12-ch DDR5-5600 ≈537 GB/s per socketr8g.8xlarge 32c/256 GB ≈$1.89/hr · r8g.16xlarge 64c/512 GB $3.77/hr · spot from ≈$0.80/hr
ModelFootprint (~Q4)Est. decode t/sWhy this pick
gpt-oss-120b (MXFP4)60.8 GB file · 64 GB+ RAM≈ 15–35The big-RAM sweet spot: 5.1B active keeps CPU decode interactive; Apache-2.0.
DeepSeek V4-FlashQ3 103 GB → Q8 162 GB≈ 4–10Frontier-adjacent (GPQA-D 88.1) with a 1M-token ctx under MIT — huge RAM is the cheapest way in.
Qwen3.5-122B-A10B~70 GB≈ 10–20Mid-size Apache MoE shaped exactly for many-core, huge-RAM boxes.
NB Dense is dead here: Llama-3.x 70B decodes at • ~4.6 t/s on 48 vCPUs — on CPU it is MoE or nothing. 512 GB also fits GLM-5.2-class Q4 (~420 GB) for best-open reasoning at batch speeds. Build llama.cpp from source for the SVE/MMLA kernels.

3MODEL CATALOG

MODELPARAMS (tot / act)ARCHTYPE · MODALITYBEST ATORGCTRYLICENSEMIN MEM ~Q4KEY BENCHMARKS & NOTES
Llama 3.2 3B3B denseTransformerinstruct · textedge chat, toolsMetaUSLlama Community (700M-MAU cap)~2 GBBest-tested Pi-class model; the largest ecosystem of fine-tunes.
Gemma 4 E2B5.1B / 2.3B eff.edge-optimizedinstruct · text+img+audon-device multimodalGoogle DeepMindUSApache-2.0~2.5 GB128K ctx; native audio input; 140+ languages.
Phi-4-mini3.8B denseTransformerinstruct · textCPU / edge reasoningMicrosoftUSMIT~2.5 GB128K ctx on a 3.8B; runs on effectively anything.
Qwen3.5-4B4B densehybrid (Gated DeltaNet)instruct+think · textsmall all-rounderAlibaba (Qwen)CNApache-2.0~2.5 GBVendor claim: 4B-class matches prior-gen 72B on many evals.
Qwen3.5-9B9B densehybrid lin-attninstruct+think · text8-GB-class defaultAlibaba (Qwen)CNApache-2.0~5 GBThe workhorse for 8–16 GB machines; ~5 GB at Q4.
Gemma 4 12B Unified12B denseTransformerinstruct · text+img+vid+audmultimodal laptopGoogle DeepMindUSApache-2.0~7 GBAdded Jun '26; MTP drafters give up to ~3× decode.
gpt-oss-20b20.9B / 3.6BMoE 32e · top-4reasoning · text+toolsagents, function callsOpenAIUSApache-2.014–16 GB≈ o3-mini quality; adjustable reasoning effort; 128K ctx.
Gemma 4 26B26B / 3.8BMoEinstruct · text+imgfast consumer / serverGoogle DeepMindUSApache-2.0~16 GBTop-10 open model on Arena (text).
Qwen3.6-27B27B densehybrid lin-attninstruct+think · textcodingAlibaba (Qwen)CNApache-2.016–17 GBSWE-bench Verified 77.2 — best open dense coder; 256K ctx.
Qwen3.6-35B-A3B35B / 3Bhybrid + MoEinstruct+think · texteveryday driverAlibaba (Qwen)CNApache-2.021–22 GBGPQA-Diamond 86.0 at ~3B active; 256K ctx.
Gemma 4 31B31B denseTransformerinstruct · text+imgopen generalistGoogle DeepMindUSApache-2.0~19 GBMMLU-Pro 85.2 · AIME'26 89.2 · #3 open on Arena.
gpt-oss-120b116.8B / 5.1BMoE 128e · top-4reasoning · text+toolsreasoning at scaleOpenAIUSApache-2.061–66 GB≈ o4-mini quality; one 80-GB GPU — or any big-RAM CPU.
Qwen3.5-122B-A10B122B / 10BMoEinstruct+think · textmid-size serverAlibaba (Qwen)CNApache-2.0~70 GBThe Apache ladder step between 35B and the 397B flagship.
DeepSeek V4-Flash284B / 13BMoE + CSA/HCA sparse attnreasoning · textlong-ctx, high value per $DeepSeekCNMIT103 GB (Q3)MMLU-Pro 86.2 · GPQA-D 88.1 · TerminalBench-2 56.9 · 1M ctx.
GLM-5.2~750B-classMoEreasoning · textopen-SOTA reasoningZhipu / Z.aiCNMIT~400+ GBGPQA-Diamond 91.2 — highest open score; server-scale only.

ALSO NOTABLE (outside these ten classes): DeepSeek V4-Pro (1.6T / 49B act, MIT — SWE-bench Verified 80.6, open-SOTA coder, ~862 GB) · Kimi K2.7 Code (~1T MoE, Modified-MIT, long-horizon coding + vision) · Kimi K3 (2.8T; weights due Jul 27 '26) · MiniMax M3 · Nemotron 3 (open weights + data + recipes) · Mistral Large 3 · Llama 4 Scout (109B / 17B act, Llama license). Benchmark scores are vendor-reported unless noted; community Q4 builds typically land slightly lower.

4RUNNING THEM

llama.cpp (GGUF) the lingua franca: CPU (AVX-512/AMX, NEON/SVE), CUDA, ROCm/HIP, SYCL, Vulkan, Metal. Every model in this report ships GGUF (stock + Unsloth dynamic quants). The 2026 flag that matters: --n-cpu-moe / --fit for MoE expert offload.

Ollama / LM Studio llama.cpp frontends; first-party libraries for gpt-oss, Gemma 4, Qwen 3.6, DeepSeek V4-Flash.

ONNX Runtime GenAI int4 ONNX builds (Phi, Gemma, Llama, Qwen small) on CPU / DirectML / CUDA; the app-embedding path.

vLLM / SGLang production serving; day-0 support for DeepSeek V4, GLM-5, Qwen 3.6; tensor/expert parallelism for multi-GPU.

OpenVINO the Intel Arc / iGPU path now that IPEX-LLM is archived.

Quant ladder: Q8_0 ≈ lossless · Q4_K_M default (≈0.55–0.6 bytes/param) · Q3/IQ3 to squeeze (V4-Flash → 103 GB) · MXFP4 is native for gpt-oss. Min-mem rule of thumb: GGUF file size + 1–3 GB KV/overhead at 8–16K ctx. Hybrid-attention Qwen 3.5/3.6 and V4's sparse attention keep long-context KV unusually small.

5CAVEATS

Decode is bandwidth-bound: t/s ≈ effective GB/s ÷ (active params × quant bytes) × 0.2–0.6 efficiency. Prefill is compute-bound and 10–100× faster on GPUs — long prompts favor GPUs far more than these decode numbers suggest.

• = measured / community-reported (sources in §6). ≈ = bandwidth-scaled estimate. Expect ±2× from llama.cpp build, context fill, and RAM config — one documented case tripled throughput just by enabling XMP.

Reasoning costs wall-clock: thinking models burn hundreds–thousands of hidden tokens; at 20 t/s a hard answer is 1–2 minutes.

Prices are a Jul-2026 snapshot taken mid-DRAM-crisis; direction is up, and the used market is the value play. AWS figures are us-east-1 on-demand.

Country column reflects the developing organization's HQ (US / CN dominate; nothing else made the cut this cycle). License ≠ jurisdiction: Apache-2.0/MIT weights are usable anywhere.

6SOURCES (ACCESSED 2026-07-24)

MODEL LANDSCAPE & SPECS

· gumloop.com — open-weight model field guide (Jul '26: DeepSeek V4, Kimi K2.7, Qwen 3.5)

· taskade.com — open-source LLM rankings + dated release timeline (GLM-5.2, Kimi K3, Qwen 3.6/3.7)

· morphllm.com — open-model benchmark tables; DeepSeek V4 / V4-Flash deep-dives

· onyx.app — open LLM leaderboard (GLM-5.2 GPQA-D 91.2; per-category leaders)

· wikipedia + ghacks + aurigait + lmstudio + layer3labs — Gemma 4 sizes, Apache-2.0 relicense, benchmarks

· codersera.com + insiderllm.com — Qwen 3.5 / 3.6 family guides, sizes, footprints, scores

· huggingface + unsloth.ai + api-docs.deepseek.com — V4 params, GGUF sizes, scores, MIT license

· mindstudio.ai + Vertex model card — gpt-oss architecture and o3-mini/o4-mini positioning

THROUGHPUT (MEASURED)

· localaimaster.com — Pi 5 12-model bench; Arc A770; AMD ROCm 7.2 guide (XTX 96 t/s)

· tinyweights + sitepoint + stratosphereips — Raspberry Pi 5/4 numbers, 17 GB/s ceiling

· runaihome.com — gpt-oss-20b consumer-GPU table (4090 225 t/s, 3060 30 t/s); Qwen3.6-35B guide

· hardware-corner.net — RTX 5090 benchmarks (32B @ 61 t/s, gpt-oss-120b @ 128K), GPU ranking, Qwen3 hardware

· insiderllm.com — Qwen3.6-35B-A3B measured: 3090 101 t/s, 3060 + expert offload 38 t/s

· techradar.com (NVIDIA/llama.cpp data) — gpt-oss-20b: 5090 282 · M3 Ultra 116 · 7900 XTX 102 t/s

· huggingface.co/ubergarm + arsturn.com — CPU-only 30B-A3B MoE: 12–15 t/s stock, 30+ tuned @ 88 GB/s

· cybernews + medium/@massimozito + carteakey.dev — gpt-oss-120b on CPUs: 5–8 t/s dual-ch DDR5, ~30 t/s @ 256 GB/s

· aws-graviton-getting-started + whileone.in + llamafile#503 — Graviton4 llama.cpp; 70B dense ~4.6 t/s

· techtactician + markaicode + propelrc — Intel Arc stacks post-IPEX-LLM archive

PRICING

· videocardprices.com — Jul '26 street: 5090 $3.68–4.5K · 5080 $1.09–1.35K · 5070 Ti $775–970 · 7900 XTX $0.8–1.1K

· techtimes + tech-insider + techpowerup — GPU memory crisis: 5090 to $4,329; RTX-50 output cut 20–40%

· alibaba electronics guide — 4090 refurb $2.5–2.8K; 4070 Ti Super $799

· raspberrypi.com + tomshardware + jeffgeerling + IEEE Spectrum — Pi price rises: 8 GB $80→$125; 16 GB $120→$205→~$300

· instances.vantage.sh + compute.doit.com + Tom's GPU tracker — r8g on-demand/spot; RX 9070 XT $649