The Week Open Weights Won

October 15th and 16th will go down as one of the biggest 48-hour stretches for open AI in years. Three U.S. organizations — NVIDIA, Meta, and the startup Reflection AI — released serious open-weight models within two days of each other, and France’s Mistral has a 1-trillion-parameter flagship whose weights are promised before the month is out. If you’ve been waiting for the “run it yourself” era of AI to actually arrive, this is what it looks like.

I covered Meta’s 816-billion-parameter Muse Spark Open last week — the one that needs a desk full of Mac Studios. This week, the story flipped: the most interesting release is the one that fits on your graphics card.

A 74-billion-parameter model that fits in 24GB

NVIDIA’s Nemotron 3.5 Bolt 74B-A9B is the headline act for anyone with a beefy desktop. It’s a 74-billion-parameter mixture-of-experts model that activates roughly 9 billion parameters per token — and thanks to native 2-bit quantization, the whole thing fits in 24GB of memory.

Quick refresher on why that matters: model weights are normally stored at 16 bits per parameter, so a 74B model needs ~150GB — datacenter territory. Quantization shrinks each weight to fewer bits after training. The catch has always been that aggressive quantization costs you capability. NVIDIA’s trick is quantization-aware training — the model trains with the precision loss simulated from the start, so it learns to be robust to it (TechTimes has a good breakdown). The result: a 74B agent model on a single 24GB card — an RTX 3090/4090/5090, or the RTX Spark laptops that went on sale last week.

Practical details: it’s on Hugging Face now, works with llama.cpp, Ollama, LM Studio, Unsloth, and vLLM, and targets local agentic workflows — coding, web research, terminal tasks. It scores 50.98 on Terminal-Bench 2.1 and 86.43 on GPQA Diamond, per NVIDIA’s own numbers (Golden Era of Super Intelligence). And it ships under the OpenMDW-1.1 license, which allows commercial use (MACCUS). A smaller sibling, Bolt E4B — a 4GB model for Jetson boards and voice assistants — lands October 23.

NVIDIA also released Super VL, a 120B-parameter multimodal sibling (12B active, 1M-token context) that reads text, images, short videos, and PDFs. That one needs a DGX Spark or H100-class hardware — datacenter fare, but open weights, datasets, and training recipes included.

Reflection’s Beam: 501B parameters, Apache 2.0, no strings

The quieter but arguably more significant release: on October 16, Reflection AI dropped the full weights of Beam under the Apache 2.0 license. That’s the genuinely permissive open-source license — no “built with” attribution requirements, no revenue-triggered side agreements, no commercial-use asterisks. Contrast that with Meta’s custom license for Muse Spark Open, and you see why licensing lawyers just perked up.

Beam is a sparse mixture-of-experts model: 501 billion parameters total, 23 billion active per token, 1M-token context, pretrained on 23.8 trillion tokens with a reinforcement-learning phase that ran on 10,500 NVIDIA GB300 GPUs for four weeks and produced over 100 million rollouts (Logeshwaran’s spec rundown). Reflection’s headline trick is token efficiency: they trained with a length penalty that rewards solving problems in fewer tokens, so the model adjusts its reasoning depth to the task — effort settings from low to max (FourWeekMBA).

The honest read on capability: Reflection’s own numbers put it at 81.6% on SWE-bench Verified and 80.1% on Terminal-Bench v2.1 — strong for coding — but independent scoring landed it at 30 on the Artificial Analysis Intelligence Index, behind GLM-5.3 (45) and Kimi K3 (44) (AI Daily Digest). Vendor benchmarks are vendor claims; treat them as a starting point.

And the homelab math is sobering: the full-precision weights are nearly a terabyte, and the NVFP4 build (~280GB) needs about four B200 GPUs (Shattered.io). This one is for serious rigs and API experimentation first — it’s already on OpenRouter at $1/$3 per million tokens — with self-hosting for the truly committed. Reflection did ship Mirror CLI, the open-source harness they trained it on, which is a nice touch.

The French giant waiting in the wings

Meanwhile, Mistral announced Large 4 (“Le Chonk”): 1 trillion parameters, roughly 50 billion active per token, trained on 3,800 NVIDIA Grace Blackwell GPUs in Europe. Weights are promised by the end of October. If you care about data sovereignty — running frontier-class weights on EU-hosted hardware with no U.S. cloud in the loop — this is the one to watch.

The closed side of the ledger

For contrast, the closed models keep charging rent: Google opened Gemini 4 Argon to developers October 13 at $2/$10 per million tokens through November (then $4/$20), and Meta’s Muse Wonder API runs $4/$20 — with the higher-scoring Wonder Full held back for Muse Code subscribers only. The open models above are explicitly positioned as the alternative.

The fine print

A few things worth keeping in mind before you download anything:

• Test 2-bit quality on your own tasks. As ODSC notes, how much capability Bolt retains at 2-bit precision on your workload is still an open question — plus it’s unclear how much extra memory long contexts need.
• Licenses are not all equal. Apache 2.0 (Beam) is the gold standard for “do what you want.” OpenMDW-1.1 (NVIDIA) allows commercial use. Meta’s Muse Model License has attribution requirements and a revenue threshold for API providers. Read before you ship.
• The cautionary note: Anthropic warned this week that small open-weight models are approaching the capability needed to self-replicate onto cheap consumer IoT hardware — Android TV boxes, not datacenters. That’s a containment risk to watch, not a reason to panic, but the “open” in open weights cuts both ways.

What I’d actually try

If you have a 24GB card: pull Bolt 74B through Ollama or LM Studio this weekend and run your usual coding/research prompts against it — that’s the experiment this whole release is begging for. If you don’t: the 4GB Bolt E4B drops October 23 and could be interesting for a Jetson or a spare laptop. And for Beam, try the API ($1/$3 on OpenRouter) before you start pricing B200s.

October 2026 might be remembered as the month open weights stopped being a hobbyist compromise and started being the default. We’ll see — but for once, the downloads are the demo.