2026年8月20日星期四

Liquid AI LFM2.5 QAD Review 2026: 4-Bit Quantization Keeps 97% Quality — Edge AI Breakthrough

Liquid AI LFM2.5 QAD Review 2026: 4-Bit Quantization Keeps 97% Quality — Edge AI Breakthrough

Liquid AI LFM2.5 QAD Review 2026: 4-Bit Quantization Keeps 97% Quality — Edge AI Breakthrough

Verdict first: Liquid AI's LFM2.5 QAD checkpoints (released August 20, 2026) are the most interesting edge-AI release of the month. By building quantization into training rather than bolting it on after, the QAD technique keeps 96.5%–97.4% of the BF16 teacher's quality in 4-bit form — the old "aggressive quantization costs visible quality" trade-off largely disappears at small scale. The 2.6B model is only ~1.3GB after quantization, runs on a phone at 213 tokens/s and a Raspberry Pi at 42 tokens/s. If you deploy LLMs on the edge, this is a genuine breakthrough worth your attention.

Liquid AI LFM2.5 QAD review 2026 — 4-bit quantization keeps 97% quality, edge AI breakthrough

LFM2.5 QAD — Liquid AI's quantization-aware distillation brings near-full precision to edge devices

What Is LFM2.5 QAD?

Released on August 20, 2026, the LFM2.5 QAD Q4_0 checkpoints cover four models — 230M, 350M, 1.2B-Instruct, and 2.6B — all published as GGUF on Hugging Face for direct use with llama.cpp. They join a series that also includes the LFM2.5-8B-A1B (BenchLM's top-rated model) and VL vision-language variants.

The positioning is clear: this is Liquid AI's play for edge and on-device inference — small models that run where big models can't.

Specifications at a Glance

ItemValue
VendorLiquid AI
Release dateAugust 20, 2026 (QAD Q4_0 checkpoints)
PositioningEdge / on-device AI deployment
Models230M · 350M · 1.2B-Instruct · 2.6B
Core techQAD — Quantization-Aware Distillation
Context window32K (230M/350M/1.2B) · 128K (2.6B)
FormatGGUF (llama.cpp compatible)
Open weightsHugging Face, $0/$0 hosted pricing
2.6B quantized size~1.3GB

The QAD Breakthrough

Traditional post-training quantization (PTQ) compresses a finished model and loses precision. QAD instead bakes the quantization constraint into training: it distills a high-precision BF16 teacher directly into a Q4_0 student. The result — quality retention across the four models:

ModelQuality retainedMatches quality of
230M97.1%Q5_K_M
350M96.5%Q5_K_M
1.2B97.4%Q4_K_M
2.6B96.6%Q4_K_M

The 4-bit checkpoints match or beat standard Q4_K_M, and match Unsloth's UD-Q4_K_XL external baseline — while being 4–33% faster than the equivalent-higher-precision quantizations at native Q4_0 memory and speed. The era where "4-bit means visibly worse" may be ending for small models.

Edge Deployment Tests (Real Hardware)

DeviceMeasured speed
Samsung Galaxy S25 Ultra (phone)213 tokens/s (230M)
Raspberry Pi 542 tokens/s (230M)
MacBook Pro (Apple M5 Max)Tested OK
NucBox (AMD Ryzen AI Max+)Tested OK

From phones to Raspberry Pi, the models genuinely run on edge hardware. Benchmarks cover reasoning, instruction-following, tool use, and agentic ability (GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, BFCLv4, plus GSM8K/AIME25 math).

Pros & Cons

Pros

  • ✅ QAD keeps 96.5–97.4% quality at 4-bit — a real breakthrough
  • ✅ 2.6B only ~1.3GB after quantization, no dedicated GPU needed
  • ✅ Runs on phones and Raspberry Pi — true edge deployment
  • ✅ Fully open GGUF weights, llama.cpp ready, $0/$0 API
  • ✅ Small models punch above their weight on data-extraction tasks

Cons

  • ❌ Small models (230M–8B) — not frontier intelligence
  • ❌ Smaller ecosystem and market influence vs big labs
  • ❌ 2.6B-Base has no inference provider yet — you self-host
  • ❌ Some figures vendor-reported, awaiting independent validation

Who Should Use It (And Who Shouldn't)

Great fit

  • Edge / on-device AI developers (mobile, IoT, embedded)
  • Teams needing local small-model deployment
  • Researchers exploring quantization distillation
  • Makers running AI on Raspberry Pi or old hardware

Poor fit

  • Users needing frontier-level intelligence (choose GPT-5.6 / Fable 5)
  • Multimodal / very-long-context workloads
  • Large-scale cloud production (limited ecosystem/providers)

Frequently Asked Questions

What is LFM2.5 QAD?

Liquid AI's edge-AI model series; the QAD Q4_0 checkpoints (230M–2.6B) released Aug 20, 2026 keep 96.5–97.4% of BF16 quality at 4-bit.

What is QAD?

Quantization-Aware Distillation — builds quantization into training, distilling a BF16 teacher into a Q4_0 student to recover most precision loss.

How big is LFM2.5-2.6B?

2.69B params, 128K context, ~1.3GB after 4-bit quantization — runs without a dedicated GPU.

Is it open source?

Yes — GGUF on Hugging Face, llama.cpp ready, no official API pricing ($0/$0), self-host at hardware cost.

How fast on edge?

213 tokens/s on Galaxy S25 Ultra, 42 tokens/s on Raspberry Pi 5 (230M model).

Is it a frontier model?

No — small models (230M–8B) for edge deployment, not flagship intelligence.

Sources

没有评论:

发表评论

SambaNova Review 2026: The Chip Company That Runs DeepSeek-R1 671B on One Rack — Faster, Cheaper, No Training

SambaNova Review 2026: The Chip Company That Runs DeepSeek-R1 671B on One Rack — Faster, Cheaper, No Training ...