←
Stephen McConnachie9,053 posts
Stephen McConnachie avatar

Stephen McConnachie

@mcnatch

Head of Data and Digital Preservation at the BFI National Archive. Working on digital preservation, audiovisual archiving, using AI with care. Opinions my own.

◷ Joined November 2010▧ 292 Photos and videos
Ihor Stepanov avatar
⟳ Stephen McConnachie retweeted
Ihor Stepanov@ihor_step·11h

🚀 10 GLiNER features you probably didn't know about! Most developers know GLiNER as a lightweight, zero-shot Named Entity Recognition model. But GLiNER has evolved far beyond traditional NER. Here are 10 powerful capabilities you might have missed 👇

Jerry Liu avatar
⟳ Stephen McConnachie retweeted
Jerry Liu@jerryjliu0·Nov 17

We’ve benchmarked over 20+ open-weight models on ParseBench-2 🔥 The leaderboard is fully public and on @huggingface. Perhaps even more importantly, it’s important to know the specifics of where open-weight models do better than others: across tables, charts, formatting, and more. The full breakdown is shown on ParseBench-2: parsebench.ai/ We’re rapidly benchmarking more models and shipping iterations on the benchmark itself. We’ve hit ~30k+ monthly downloads in the past few days. Hugging Face: huggingface.co/datasets/llam… Leaderboard: parsebench.ai/

Post attachment
merve avatar
⟳ Stephen McConnachie retweeted
merve@mervenoyann·Nov 17

it's a skill issue to use frontier LLMs to detect objects or do vision instead, if you use LLMs as vision interns you can train a detector for $6 that runs at little cost. but your agent is a bad vision intern, it only knows old models that are proprietary and makes rookie mistakes so I built something for it, I basically distilled myself for you to use🤝 I put my favorite models from keypoint detection to OCR as a tool for your agent to call and I also made an e2e object detection trainer (from labelling to VLM as a judge to training and eval) if you're interested in it and the results, my talk at @aiDotEngineer WF is finally out, you can watch it here youtube.com/watch?v=dKcTBQzR… many thanks @swyx on the next post you can find the github repository, I also shipped decision model support and updated VLMs (LFM3, Qwen3.6/8)

Vik Paruchuri avatar
⟳ Stephen McConnachie retweeted
Vik Paruchuri@VikParuchuri·Nov 16

We shipped Chandra 2.7! Nothing massive, just steady OCR improvement: - Better tables: balanced 91.6 → 94.2 on OmniParseBench - Better on non-English and rotated pages

Post attachment
merve avatar
⟳ Stephen McConnachie retweeted
merve@mervenoyann·Nov 13

today, llama.cpp releases v1 🚀 llama.cpp team commits to keep covering every platform and every model, while keeping the platform robust and lean for everyone that builds on it 🫡 read more about the release here huggingface.co/blog/ggml-org…

0:33
Daniel van Strien avatar
⟳ Stephen McConnachie retweeted
Daniel van Strien@vanstriendaniel·Nov 13

OCR models keep getting cheaper! I wanted to see how much, and where the savings come from. An agent helped me run 8 open OCR models from the last 18 months on the same 850 pages, on one GPU on @huggingface Jobs. - $0.51 to $0.095 per 1,000 pages. A straight line through the eight halves about every 10 months, but most of the drop is one new model: LightOnOCR-3-0.8B from @LightOnIO. - How you run a model changes the cost too. GLM-OCR on 250 dense pages: $1.57 per 1,000 run the way I ran it in February, $0.40 today (many pages in flight, pages fitted to 2,048 px, newer @vllm_project, bigger GPU). - Full-size scans did not help. For three models, they cost 1.3 to 4.2x more than 2,048 px, with no accuracy gain on book pages. - On 282 checked book pages, every model gets its median page over 99% right. Two of the cheaper ones make about 1.5x the errors of the oldest. - On dense two-column 1771 print, the oldest model was preferred over most of the newer ones. The newest about ties with it. So still worth testing a few models on your own pages. Comparing all eight on 850 pages cost about $6 of GPU time. Writing a full post on this!

Post attachment
Daniel van Strien avatar
⟳ Stephen McConnachie retweeted
Daniel van Strien@vanstriendaniel·Nov 7

Recent text-to-speech models small enough to run on a CPU: phonon-31m-en - 31M Paradee-8M - 7.9M Leokani-tts - 7M (4.5M version too) PiVo - 1.4M (vocoder included) Paradee runs ~62x real time on a single CPU thread.

Daniel van Strien avatar
⟳ Stephen McConnachie retweeted
Daniel van Strien@vanstriendaniel·Nov 6

Nemotron 3.5 Transcribe from @NVIDIAAI is very fast with @vllm_project! 11.3 hours of 1971 Nixon tapes on one RTX PRO 6000 (NVFP4): under 3 minutes of transcription. Whole Job, setup included: about $0.30. For batch jobs, I run vLLM inside the @huggingface Job and send many requests at once using my tool saturate, which tunes the concurrency as it goes. A 79.7-hour run took just under 10 minutes. Recipe: huggingface.co/datasets/uv-s…

Post attachment
Daniel van Strien avatar
⟳ Stephen McConnachie retweeted
Daniel van Strien@vanstriendaniel·Nov 3

OCR models often get stuck in loops, so I asked Claude Opus if it could stop one from looping by editing its weights. It changed 0.0085% of them and with no training, it now reads this 1881 page instead of writing 16,399 characters of dots! Prompt was basically: "read this, find a smallish OCR model and fix its repetition doom loops. You can spend up to 5 hours and $100 on @huggingface Jobs." It only took under 3 hours in the end and about $10 of Jobs compute.

Post attachment

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…

Post attachment
Daniel van Strien avatar
⟳ Stephen McConnachie retweeted
Daniel van Strien@vanstriendaniel·Oct 30

Tried this on a bunch of old newspapers. - 1760s Fraktur, polytonic Greek, pre-reform Russian: all read well to my untrained eye; layout included - Big broadsheets need full-res input: error rate 25% → under 10% (these are examples where many OCR models struggle.

Post attachmentPost attachmentPost attachment

Introducing Surya OCR 2.5! - Most accurate open OCR model under 1B params (85.9 on OmniParseBench) - 1.7x faster than Surya 2, 4.5 pages/s on H100 - 689M params, with character tokenization

0:28
Xiaomi Tech avatar
⟳ Stephen McConnachie retweeted
Xiaomi Tech@XiaomiTech_·Oct 30

From recognizing text to understanding documents. 📄 Meet Xiaomi-OCR-0, an open-source document understanding model with just 0.8B parameters. It brings document parsing, key information extraction, and OCR-oriented visual question answering into one lightweight model. It scores 95.24 on Real5-OmniDocBench and 83.2 on average across five OCR-oriented VQA benchmarks. On OmniDocBench v1.6, Xiaomi-OCR-0 ranks first in document parsing performance among models with fewer than 1B parameters. Model weights, inference code, an interactive demo, MCP service, and Agent Skill are now open source. 👇 🤗 Huggingface: huggingface.co/xiaomi-resear… 🔗 GitHub: github.com/xiaomi-research/X… 📄 Demo: huggingface.co/spaces/loongw… #XiaomiAI #OCR #OpenSource

1:07
Doğukan avatar
⟳ Stephen McConnachie retweeted
Doğukan@DogukanUrker·Oct 29

running NVIDIA Nemotron 3.5 Transcribe (30B-A4B) on a single RTX 3060. 12GB VRAM, 16GB RAM. (config below) 90 seconds of three people talking over each other, done in 27s. speaker labels and word timestamps in one pass. 1 hour context. His Girl Friday (1940) is famous for the fastest overlapping dialogue ever filmed, so i threw it at it. stock NeMo-Speech.cpp can’t offload experts to cpu yet, so i added a small patch: n_cpu_moe (like llama.cpp --n-cpu-moe) plus a higher output budget, because the default cuts dense crosstalk off halfway. config: nemo-speech serve --asr-model NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-AD-IQ4_NL.gguf --asr.salm.n_cpu_moe 40 --asr.salm.max_sequences 1 --asr.salm.n_ctx_per_sequence 73728 --asr.salm.max_context 73728 --asr.salm.n_batch 1024 --asr.salm.encoder_max_batch 4 env: NEMO_SALM_LOAD_MODE=mmap NEMO_SALM_OUT_TPS=12

1:30

Meet NVIDIA Nemotron 3.5 Transcribe 🎙️ An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint. Watch it keep track of who said what when everyone’s talking at once. 1/ 🧵

2:00
Daniel van Strien avatar
⟳ Stephen McConnachie retweeted
Daniel van Strien@vanstriendaniel·Oct 29

Added a script to run this on @huggingface Jobs with one command (@vllm_project + NVIDIA's NeMo plugin, inside the job). It's very fast: 4.9 hours of 1940s radio transcribed in 2.5 min once the model is up. Whole job, setup included: $0.48. Speaker tags + word timestamps, here on Hitchcock's 1940 Suspense audition: [5:21] spk2: No, I'll go. [5:39] spk1: Is it not true that you let lodgings? huggingface.co/datasets/uv-s…

Meet NVIDIA Nemotron 3.5 Transcribe 🎙️ An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint. Watch it keep track of who said what when everyone’s talking at once. 1/ 🧵

2:00
Ashley Gold avatar
⟳ Stephen McConnachie retweeted
Ashley Gold@ashleyrgold·Oct 28

NEW from me: Britain’s AI testing agency outguns US Britain wants to pour another £120 million — roughly $160 million — into a government lab that tests the world’s most powerful AI models, an infusion of cash that alone is more than 10 times the annual budget of its US counterpart. It contrasts sharply with the approach in the US, where federal efforts to evaluate AI are receiving far less cash and are shared by different parts of government. semafor.com/article/10/27/20…

NVIDIA AI avatar
⟳ Stephen McConnachie retweeted
NVIDIA AI@NVIDIAAI·Oct 28

Meet NVIDIA Nemotron 3.5 Transcribe 🎙️ An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint. Watch it keep track of who said what when everyone’s talking at once. 1/ 🧵

2:00