"look at the data" has to be the highest alpha advice for working on LLMs; I'm currently working on optimizing LFM VL models and I realized I could improve my intuitions on vision datasets; a thread where I go through the popular vision benchmarks, 1/x

Top Models For your Hardware 2026 -- 8GB -- - lfm-2.6B: very solid, trained on tons of data, obviously limited by good place to start Use this for snippet tool calls, or to play around and build infra -- 16GB -- I am a big fan of Ornith models, they do well at tool calling and they certainly push models farther Gemma is a more well rounded model for chat, and vision IMO -- 24GB up to 96GB -- Qwen3.8-27B is IMO the first model here you can code with, it's really strong. Use the exl3 versions. -- 96GB up to 196GB -- Qwen3.8-Flash-Next is where it's at. Phenomenal model, very fast even on slower hardware, kv cache is smaller and part of the model is basic enough to be offloaded to ram -- 196GB up to 384GB -- GLM-5.3-Flash is frontier at home, built to run on 10,000$ of hardware. It's really a gift. It is natively multimodal, image/video/ audio? It hold up over 1 million tokens in context, and is hybrid attention, it holds on higher concurrency. Very good for 3D, coding, hacking, art. -- 384GB up to 512GB -- Sacrifice speed for intelligence, GLM-5.3 is tied for #1# Open Weight model, and is the best for pure coding and systems work. I told you August was going to be awesome, now we go to slower months. ------------- It's not Local AI we have to worry about anymore.

catching up on some bookmarked reads from the summer, reading Antidoom from @liquidai small reasoning models get stuck more easily when the task involves a long thinking trace and a hard problem. It starts repeating the same word over and over again ("Wait", "Alternatively"…), each repetition makes the next one likelier, and the generation is spent before it reaches an answer they measured it, 10.2% of completions for an early LFM2.5-2.6B checkpoint and 22.9% for Qwen3.5-4B at greedy. After training those drop to 1.4% and 1.0% the fix is FTPO (final token preference optimization). What I like is how narrow it is, it only touches the single token where the loop starts three ways it differs from DPO: > trains one token position, mid-generation, instead of whole sequences > spreads probability across ~20 plausible alternatives instead of swapping one overtrained token for another > keeps the regularizer in logit space, no softmax, so the rest of the vocabulary stays put the third one is what makes it usable. If you want to edit one position without disturbing the model, you can't have a loss that reshuffles the other 150k logits on the way and their explanation abt the result: the training teaches the model nothing new about math or code, it clears the failure mode that was blocking answers the model could already produce full blog > FTPO itself comes from Antislop, where it was built to strip overused phrasing. LiquidAI retargeted it to doom loops and under the hood it's a subclass of TRL's DPOTrainer with compute_loss overridden, around 90 lines of loss and no new trainer we documented that pattern in TRL's docs

Which model should you run on the iPhone 17 Pro? Announcing our new intelligence and inference testing for small models on mobile devices: independent measurement of how capable small models are in typical on-device tasks, and how they perform on popular phones - in partnership with @liquidai We have partnered with @liquidai to deliver mobile device inference benchmarking, covering a range of models in 4-bit or lower precision on the iPhone 17 Pro and Galaxy S26 Ultra We’re publishing our phone-scale intelligence evaluation results in combination with @liquidai's inference performance benchmarks to give users and developers a holistic view of how small models are performing on phones Inference benchmarking is conducted in a controlled environment using Liquid AI’s inference benchmarking software, which Artificial Analysis has examined and is open-sourced on Liquid AI’s GitHub. Results cover end-to-end generation time, output speed, peak memory usage and other metrics. The inference benchmarking app, ‘Pipette’, is available to download for free on iOS and Android, allowing users to test a variety of models on their own devices We are ranking phone-scale model intelligence based on each model’s average score in five evaluations chosen for the task-based work these models do in practice: BFCL, IFBench, AA-Omniscience, GPQA Diamond and MATH-500. These evaluations are run by Artificial Analysis using our independent methodology. By default, we limit models to 16K context on each of these evaluations, representing the lack of memory space for significant KV cache on mobile devices. This leads to some intelligent but verbose models dipping in relative score - they were not designed for the constraints that phone memory imposes on token use We are defining our portable device category as including models that fit within 8 GB of memory after quantization, including KV cache, at 8K context We expect both the intelligence and inference benchmarks to evolve over time, as new models, devices, inference frameworks and quantization techniques are released. Our pages will remain up to date with these new additions Initial results: ➤ Nanbeige4.2-3B and LFM2.5-2.6B share the top average evaluation score at 63 (with a 16K context limit), ahead of Ornith-1.0-9B at 62 and Qwen3.5 9B (Reasoning) at 61. LFM2.5-2.6B achieves its score more efficiently: on an iPhone 17 Pro it answers a standard 1,024-token prompt in 8.0s using 2.3 GB of memory, against 21.4s and 4.0 GB for Nanbeige4.2-3B and 25+ seconds and 6.9 GB for the two 9B models ➤ The 16K context limit shapes the leaderboard: as an example, Qwen3.5 9B (Reasoning) spends 74.5M output tokens across one pass of the benchmark set, hitting the 16K limit on 29% of its generations and landing in fourth place overall. With the limit raised to 64K, Ling 3.0 Tiny takes first place with a score of 66, ahead of Nanbeige4.2-3B (65), Qwen3.5 9B (Reasoning, 64), and LFM2.5 2.6B (64). But a 64K window does not fit in mobile phone memory, and at 55 output tokens/s on an iPhone, generating 64K tokens could mean a 20+ minute wait and a lot of battery use. This is why our primary results are capped at 16K, but we're also publishing a set of results capped at 64K, and another capped at one minute of generation time ➤ The speed-intelligence Pareto frontier is short: six models are unbeaten on both intelligence and speed on an iPhone 17 Pro: LFM2.5-230M (27 at 0.9s), MiniCPM5-1B (45 at 2.9s), LFM2.5-8B-A1B (58 at 5.7s), Ling 3.0 Tiny (59 at 5.7s), LFM2.5-2.6B (63 at 8.0s) and Nanbeige4.2-3B (63 at 21.4s). LFM2.5-8B-A1B and Ling 3.0 Tiny are mixture-of-experts models that activate ~1B parameters per token, which is how they answer in under 6s with 8B-class weights ➤ Leading models have opposite strengths: Nanbeige4.2-3B is the most balanced (76% on BFCL, 96% on MATH-500, 67% on GPQA Diamond); Qwen3.5 9B (Non-reasoning) is the strongest tool caller (77% on BFCL) and scientific reasoner (79% on GPQA Diamond); LFM2.5-2.6B follows instructions best of any model measured on the iPhone (59% on IFBench), clears 90% on MATH-500 and hallucinates far less on AA-Omniscience (79% non-hallucination, against 33% for Nanbeige4.2-3B and 24% for Qwen3.5 9B (Reasoning)) More details below in thread ⬇️

Today, we release DSpark draft models for LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. These add a speculative decoding path that trades a minimal memory increase for a large decoding speedup without changing output quality. A lightweight draft model proposes a block of candidate tokens and the target model verifies them in a single forward pass. Across MATH500, GSM8K, HumanEval, MBPP, and MT-Bench at batch size 1: > Up to 3.18x throughput on an H100: LFM2.5-8B-A1B on MATH500, 428 → 1362 tok/s > Up to 2.87x on an M4 Max MacBook Pro: LFM2.5-1.2B-Instruct on HumanEval, 136 → 389 tok/s > LFM2.5-2.6B means: 2.67x on the H100 (323 → 864 tok/s), 2.27x on device (61 → 139 tok/s) > Under greedy decoding, the emitted sequence is identical to baseline by construction, so benchmark accuracy is unchanged. Each draft model is around 300M parameters, with embedding and LM head tied to its target model. The gain shows most in agentic workloads, where the model reasons before every tool call and the user waits through it all: on BFCL multi-tool scenarios, DSpark cuts LFM2.5-2.6B latency by nearly 50% on average. 🧵

So you’re telling me I can swap my local LFM2.5-2.6B from F16 to QAD Q4_0 and go from: 5.4 GB → 1.6 GB 21 → 64 tok/s 3.0s → 1.2s tool-call latency while keeping ~97% of BF16 performance? @LiquidAI what did you just do 😭