aliteq.

Type to search every guide, tool and GPU page.

How much VRAM Tencent Hy4 Flash (MoE) needs, and what it costs to run

Some links on this page are referral links: we may earn a commission, at no cost to you.

Trying to decide what to buy? The best GPU for Tencent Hy4 Flash (MoE) ↓ · Got a card already? See what your PC can run →

220.3B paramsMoE · 512 experts, 16 active1024k max contexttencent-hy-license

Cheapest way to rent it right now

$0.24/hr

NVIDIA DGX Spark (GB10 Grace Blackwell) · Q4_K_M · on-demand · ≈ $58/mo at 8h/day

Live price captured Thu, 19 Nov 2026 17:20:08 GMT.It never changes which card is cheapest.

Memory needed, by quantisation

At 16k context. Weights + KV cache + ~0.8 GB overhead. Memory is in binary GB (1,024³ bytes), the unit GPU memory is sold in; Hugging Face and Ollama list downloads in decimal GB, about 7% more.

QuantWeightsKV cacheTotalQuality
F16410.2 GB0.2 GB411.2 GBFull precision. Reference quality, twice the size of Q8 for no practical gain in most chat use.
Q8_0217.9 GB0.2 GB218.9 GBEffectively lossless. Use when VRAM is not the constraint.
Q6_K168.2 GB0.2 GB169.2 GBVery close to Q8 at meaningfully less memory. A safe high-quality pick.
Q5_K_M145.6 GB0.2 GB146.6 GBSmall, hard-to-notice quality loss. Good balance.
Q5_0142.0 GB0.2 GB143.0 GBOlder-style 5-bit. Q5_K_M is usually the better pick at the same size.
Q4_K_M124.4 GB0.2 GB125.3 GBThe community default. Best quality-per-gigabyte for most people.
Q4_0116.7 GB0.2 GB117.7 GBOlder-style 4-bit, measurably worse than Q4_K_M at a similar size. Avoid unless required.
IQ4_XS109.0 GB0.2 GB110.0 GBNewer 4-bit, smaller than Q4_K_M with comparable quality. Needs a recent llama.cpp.
Q3_K_M100.3 GB0.2 GB101.2 GBNoticeable degradation. Use to fit a larger model that would otherwise not run.
Q2_K85.9 GB0.2 GB86.9 GBHeavy degradation. Almost always better to run a smaller model at Q4_K_M instead.
IQ3_XS84.6 GB0.2 GB85.6 GBAggressive. Usually better than a smaller model at Q4, but test before trusting it.

Cards that can run it

Best quantisation each card fits at 16k context, with live rental price where we track one.

GPUVRAMBest fitUses~tok/sRent from
Apple M3 Ultra512 GBF16411 GB——
Apple M5 Ultra512 GBF16411 GB——
Apple M2 Ultra192 GBQ6_K169 GB——
Apple M5 Max (18-core CPU / 40-core GPU)128 GBQ4_K_Mtight125 GB——
NVIDIA DGX Spark (GB10 Grace Blackwell)128 GBQ4_K_Mtight125 GB—$0.240/hrVast.ai
Apple M4 Max (16-core CPU / 40-core GPU)128 GBQ4_K_Mtight125 GB——
AMD Ryzen AI Max+ 395 (Strix Halo, 128 GB)128 GBQ4_K_Mtight125 GB——
NVIDIA RTX PRO 6000 Blackwell Workstation Edition96 GBQ2_Ktight87 GB—$0.467/hrVast.ai
Apple M2 Max96 GBQ2_Ktight87 GB——
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition96 GBQ2_Ktight87 GB—$0.333/hrVast.ai

The best GPU for Tencent Hy4 Flash (MoE)

Ranked from live rental prices and computed VRAM fit, not opinion. Every card is judged at Q4_K_M with a 8K context, so the comparison is fair; a card that could only run a crushed version isn't listed.

Best value

—

Best throughput per dollar among cards that fit.

Cheapest that runs it

—

Lowest-cost card that still fits the model.

No compromise

—

Most headroom and highest throughput.

The value column ranks renting, because that's what we can price precisely (prices captured Thu, 19 Nov 2026 17:20:08 UTC). Buying a card? The fit and throughput columns are what you need; we don't publish a purchase-price ranking because we don't hold verified street prices. Tokens/sec is an estimate, not a benchmark. Got a card already? See everything your PC can run.

hy4-flash

How these numbers are produced

  • Model shape comes from the model's own config.json — 12 layers, 32 attention heads, 1 KV heads.
  • Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
  • KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
  • Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
  • This is a mixture-of-experts model. Generation reads only the experts active for each token (16 of 512), so throughput is scored on ~11.3B active parameters, not the full 220B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Hybrid attention, 42 layers, all MLA (kv_lora_rank 512 + qk_rope_head_dim 64 = 576 cached numbers per token per layer, mapped as kv_heads 1 x head_dim 288 x 2 = 576, same convention as GLM-5.3): 12 sparse-attention (IndexPool/DSA) layers hold the full-context latent cache, 30 sliding-window layers keep only a 512-token window, so `layers` = 12; the window layers add a small fixed amount and the compressed indexer keys add a little on top. Card: 220B total = 155.5B backbone + 63.5B Engram n-gram tables (hash lookups, still have to be stored) + 0.43B ViT + 0.85B DFly2 drafter; 11.3B text-active per token (+0.43B vision), headline 12B. 512 routed experts top-16 + 1 shared, latent dim 1,536; layer 0 dense (13,824). Up to 1M context, text + image in. Released BF16 (an FP8 copy is tencent/Hy4-flash-FP8), so the quant ladder starts from BF16. params = HF safetensors total. HF config.json + card read Nov 21, 2026.
  • Some memory-bandwidth figures are not yet independently verified.Where that's the case we show no tokens/sec at all rather than a number we can't stand behind.
  • Rental prices are pulled hourly from provider APIs (last updated Thu, 19 Nov 2026 17:20:08 GMT). Spot/interruptible pricing can change or vanish without notice.

Try it interactively

Drag the context slider and watch the KV cache fill the card

Running Tencent Hy4 Flash (MoE): common questions

How much VRAM do you need to run Tencent Hy4 Flash (MoE)?
At Q4_K_M and at 8k context, Tencent Hy4 Flash (MoE) needs about 125.2 GB — 124.4 GB of weights plus 0.1 GB of KV cache and ~0.8 GB overhead. So you want a card with at least that much free memory; the KV cache grows if you use longer context.
Can an RTX 4090 run Tencent Hy4 Flash (MoE)?
Not at Q4_K_M — Tencent Hy4 Flash (MoE) needs more than the card's 24 GB at that quality and 8k context. You'd have to drop to a lower quantisation (worse quality), shorten the context, or use a bigger card or two cards.
Can an RTX 3090 run Tencent Hy4 Flash (MoE)?
Not at Q4_K_M — Tencent Hy4 Flash (MoE) needs more than the card's 24 GB at that quality and 8k context. You'd have to drop to a lower quantisation (worse quality), shorten the context, or use a bigger card or two cards.
Can an RTX 4060 Ti 16GB run Tencent Hy4 Flash (MoE)?
Not at Q4_K_M — Tencent Hy4 Flash (MoE) needs more than the card's 16 GB at that quality and 8k context. You'd have to drop to a lower quantisation (worse quality), shorten the context, or use a bigger card or two cards.

Worth reading before you buy

Other models