PrismML

PrismML (also styled Prism ML; legal name Prism ML, Inc.) is an American artificial intelligence company, based in Pasadena, California, that publishes extremely low-bit versions of open-weight models under the name Bonsai.^[1][16] Its models store almost every weight of the network as either a sign bit (1-bit, values {-1, +1}) or a ternary digit (values {-1, 0, +1}, about 1.58 bits of information), each group of 128 weights sharing one 16-bit scale factor.^[19][21][22] The company came out of stealth on 31 March 2026 with 1-bit Bonsai 8B and grew out of research by Caltech electrical engineering professor Babak Hassibi, its co-founder and chief executive.^[1][3][4] As of 30 October 2026 its flagship is Ternary Bonsai 2 27B, a 5.95 GB ternary build derived from Qwen3.8-27B that PrismML says keeps 98.2% of the original model's benchmark average.^[8][19]

All Bonsai weights are released on Hugging Face under the Apache License 2.0.^[1][19] Several of the packing formats need PrismML's own kernels, distributed through forks of llama.cpp and Apple's MLX, although parts of that support have since been merged into mainline llama.cpp.^[19][30][31] The benchmark results in PrismML's announcements come from the company's own evaluations.^[8][20]

History

Stealth exit and the 1-bit family (March 2026)

PrismML announced itself on 31 March 2026 with a press release dated from Pasadena. It said the company was introducing "the world's first commercially viable 1-bit large language models, built on groundbreaking research developed at Caltech", led by 1-bit Bonsai 8B, and released 1-bit Bonsai 4B and 1.7B alongside it.^[1] Hassibi was quoted saying the team "spent years developing the mathematical theory required to compress a neural network without losing its reasoning capabilities".^[1] The company said the 8B model had been trained using Google v4 TPUs.^[1]

The Register covered the launch on 4 April 2026. It reported that PrismML's approach is based on work by Hassibi and colleagues at Caltech, and placed it in a longer line of low-bit research that includes the 2024 paper "The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits".^[3] On 1 April 2026 a pull request written on behalf of the Bonsai team ("We just released Bonsai 1-bit models") proposed adding the Q1_0 format to mainline llama.cpp; it was merged on 6 April 2026.^[30]

Ternary and image models (April to May 2026)

On 16 April 2026 PrismML announced the Ternary Bonsai family (8B, 4B and 1.7B), which uses three weight values instead of two. The company said Ternary Bonsai 8B averaged 5 benchmark points more than 1-bit Bonsai 8B while needing about 600 MB more memory.^[9][10] On 26 May 2026 it released Bonsai Image 4B, 1-bit and ternary versions of the FLUX.2 Klein 4B image model (part of the FLUX.2 family), together with an iOS app called Bonsai Studio.^[11][12]

Bonsai 27B and Apple (July 2026)

On 14 July 2026 PrismML released Bonsai 27B, derived from Qwen3.6-27B, in a 1-bit variant of about 3.9 GB and a ternary variant.^[13][14] PrismML's press release said the 1-bit variant ran at 11 tokens per second on an iPhone 17 Pro; its model card gives the figure for an iPhone 17 Pro Max.^[13][23] The same day CNBC reported that Apple was in talks with the company. Hassibi told CNBC that Apple and other companies had been evaluating PrismML's models for speed, energy efficiency and on-device performance; he called the discussions very early but said "things are progressing nicely". Apple did not immediately respond to CNBC's request for comment.^[5][34] The Next Web added that The Information had first reported Apple's interest the week before.^[34] 9to5Mac wrote that it was "clear" PrismML was "touting communication with Apple as a way to generate buzz" around the release.^[33] Hassibi told CNBC that Google's Gemma models were next in the pipeline, followed by much larger models.^[5] Summarising the accompanying whitepaper, The Decoder reported a battery test of roughly 672 generated tokens per percentage point of charge for the 1-bit model on an iPhone 17 Pro Max.^[32]

Bonsai 2 (September to October 2026)

PrismML released Ternary Bonsai 2 27B, based on Qwen3.8-27B, on 17 September 2026.^[8][15][35] TechCrunch's coverage that day described PrismML as "founded by a group of Caltech researchers" and quoted Hassibi saying the company's next models, "hopefully in the next couple of months, will be in the several-hundred-billion-parameter range".^[6] At Qualcomm's Snapdragon Summit on 23 September 2026, a 2-billion-parameter vision-language model built on Bonsai 1.7B was shown running on smart glasses that use the Snapdragon AR1 Gen 1 platform.^[16][7] TechCrunch noted that no smart glasses running PrismML's model had been announced.^[7] On 15 October 2026 the company added Ternary Bonsai 2 2B, 4B and 9B, based on Qwen3.5.^[17][18]

People and funding

The company's "Meet the team" page lists four co-founders:^[4]

The same page names Ion Stoica, Buno Pati and Julie Schoenfeld as advisors.^[4] TechCrunch describes Stoica, a co-founder of Databricks and director of UC Berkeley's Sky Computing Lab, as an adviser to the company.^[6]

Hassibi is a professor of electrical engineering at Caltech.^[3] He is the first author, with David G. Stork, of "Second Order Derivatives for Network Pruning: Optimal Brain Surgeon" (NIPS 1992), an early paper on removing unimportant weights from trained neural networks (see pruning).^[37] CNBC reported that the Bonsai technology emerged from Hassibi's research group and that Caltech owns the underlying patents and licenses them exclusively to PrismML.^[5]

Reported funding figures differ by date. CNBC and The Next Web reported in July 2026 that PrismML raised a $16.25 million seed round in March, backed by Khosla Ventures and other investors.^[5][34] TechCrunch described a $22.25 million seed round in September 2026.^[6] PrismML's March press release said the company was backed by Khosla Ventures and Cerberus Ventures, with compute grants from Google and Caltech.^[1] Its later press releases say it "was founded with support from Khosla Ventures, Cerberus, and Google, with continuing support from Samsung"; the 23 September release adds Gates Frontier and Jeff Dean to that list.^[8][16]

Technology

Weight formats

Every Bonsai language model uses group-wise scaling: weights are split into groups of 128, and each group stores one FP16 scale. In the 1-bit format a stored 0 means minus the scale and a stored 1 means plus the scale, which costs 1 bit plus 16/128 bits, or 1.125 bits per weight.^[21] In the ternary format each weight is minus the scale, zero, or plus the scale. A ternary value carries log2(3), about 1.585 bits of information, so the effective cost is about 1.71 bits per weight once the scale is included.^[14][19]

PrismML stresses that the low-bit representation covers the embeddings, attention projections, MLP projections and the output (LM) head, with "no higher-precision escape hatches".^[2][10] In Bonsai 2 27B a small set of tensors (the recurrent-state path of the linear-attention layers and the normalization weights, 26.2 million parameters or 0.0976% of the language model) is kept in higher precision, which raises the whole-model figure to 1.72 bits per weight.^[19][20] Bonsai 2 also stores weights in a rotated basis: each matrix is transformed blockwise by a fixed Walsh-Hadamard rotation (block size 1024, with fixed sign flips) before the ternary values are assigned, and the runtime applies the matching transform to activations. The whitepaper cites SpinQuant for this idea.^[19][20]

Bonsai models keep the architecture of the model they are derived from. The model cards say Bonsai 2 27B is "derived from Qwen3.8-27B" with the "architecture unchanged", and the 2B, 4B and 9B cards say the same of their Qwen3.5 bases.^[19][25] The cards and whitepapers document the stored formats and the inference kernels; PrismML calls the underlying design proprietary, and its press releases say the 27B and Bonsai 2 models were trained on Google v5 TPUs.^{[2][9][8][13][17]}

Relation to BitNet and conventional quantization

The {-1, 0, +1} alphabet is the same one used by Microsoft's BitNet b1.58, whose authors present ternary weights as "a new scaling law and recipe for training new generations of LLMs" (see BitNet and quantization-aware training).^[39] The Register cited that paper as prior work when covering PrismML's launch.^[3] Bonsai instead starts from existing open-weight checkpoints from Alibaba's Qwen family and, for images, FLUX.2 Klein 4B.^[12][19][21] When PrismML proposed a 2-bit ternary format for llama.cpp, it explained why it did not reuse llama.cpp's existing ternary types TQ1_0 and TQ2_0: those use groups of 256 weights rather than 128 and were CPU-only and harder to accelerate on Metal and CUDA.^[31]

PrismML compares Bonsai against conventional post-training quantization builds of the same base model, such as the IQ2_XXS and UD-Q4_K_XL GGUF files.^[19] The Bonsai 27B model card argues that conventional labels understate real bit-widths, citing a "2-bit" build of Qwen3.6-27B that averages 2.8 bits per weight at 9.4 GB.^[23] The company also says conventional low-bit builds fail unevenly: in its whitepaper's words, "long context reasoning and coding deteriorate first, while surface fluency and short tool calls retain high performance".^[19][20]

Intelligence density

PrismML promotes a metric it calls intelligence density: the negative logarithm of a model's average error rate on a benchmark suite, divided by the model's size in gigabytes.^[2][3] The Register described it as "a metric that shows its models in a good light".^[3] The definition is not applied uniformly: the 1-bit Bonsai 8B card uses the natural logarithm, while the Bonsai 2 27B card and whitepaper use log base 2.^[19][20][21]

Models

The table lists PrismML's published models with the base model and sizes given on each model card or announcement. Release dates are announcement dates; some Hugging Face repositories were created days or weeks earlier.^[28]

The ternary 8B family's announcement gives a smaller footprint (1.75 GB for the 8B) than the packed Q2_0 file listed on its model card (2.18 GB), because the 2-bit packing stores each three-valued weight in a 2-bit slot.^[10][22] Bonsai 27B and later models accept images as well as text; the vision tower ships as a separate optional file.^[14][19] For the smart-glasses demonstration PrismML built a separate 2B vision-language model from Bonsai 1.7B and a 4-bit vision encoder.^[16]

Inference software

PrismML maintains forks of llama.cpp (default branch prism), MLX and mlx-swift, plus a Bonsai-demo repository that its model cards call "the source of truth for running these models".^[19][29] Much of the low-bit support has since moved into mainline llama.cpp. The Q1_0 format (1-bit, groups of 128) was merged for CPU on 6 April 2026, and a CUDA backend followed on 15 April.^[30][38] A Q2_0 ternary format was merged in July 2026. At the maintainers' request it uses groups of 64, so PrismML keeps a group-of-128 variant called PQ2_0 in its fork.^[31]

The Bonsai 2 27B files still need the fork. Its model card says "Stock llama.cpp will not run these files": mainline rejects PQ2_0 and PTQ1_0 as unknown types.^[19] The repository ships no Q2_0 file, and its known-issues page says the F16 file loads in stock llama.cpp but produces garbled output because it depends on metadata that only the PrismML build applies. PrismML's October cards for the smaller Bonsai 2 models say recent mainline builds include the Hadamard activation runtime and can run their Q2_0 files.^[19][25] PTQ1_0 packs ternary digits densely (1.75 bits per weight, 5.95 GB), while PQ2_0 gives each digit a 2-bit slot (2.13 bits per weight, 7.21 GB). PrismML reports that neither is faster everywhere: for decoding, PTQ1_0 wins on Ada-generation GPUs and the L4, and PQ2_0 wins on H100, A100 and Blackwell cards.^[19]

On GitHub, the Bonsai-demo repository had 3,467 stars and the llama.cpp fork 986 stars as of 30 October 2026.^[29]

Bonsai 2 27B

Ternary Bonsai 2 27B has 27.36 billion parameters: a 24.35-billion-parameter language backbone of 64 blocks, 2.54 billion in the embedding and LM head, and a 0.46-billion-parameter vision tower. Like its base model, it uses hybrid attention (about 75% linear attention and 25% full attention) and a 262K-token context window.^[19] The FP16 original is about 54 GB.^[19]

Vendor benchmarks

PrismML reports two different benchmark suites for the model, both run in thinking mode with EvalScope and vLLM on NVIDIA H100 GPUs:^[19][20]

The 14-benchmark card suite covers MMLU-Redux, MuSR, GSM8K, MATH-500, AIME25, AIME26, HumanEval+, MBPP+, LiveCodeBench, IFEval, IFBench, BFCL v3, MMMU-Pro and OCR Bench v2. On it the conventional UD-Q4_K_XL build (17.6 GB) scores 85.18.^[19] The 20-benchmark whitepaper suite drops MuSR and MMMU-Pro and adds GPQA Diamond, AA-LCR, BigCodeBench, τ²-Bench, CharXiv, A-OKVQA, OmniDocBench v1.6 and RealWorldQA.^[20] In the whitepaper PrismML also reports 52.8 on Terminal-Bench 2.1 and 60.8 on SWE-bench Verified, against 69.7 and 80.6 for the full-precision Qwen3.8-27B, or roughly three quarters of full-precision performance.^[20] By category, the card shows the largest losses in knowledge and reasoning (79.86 against 85.55) and vision (66.19 against 71.36).^[19]

The vendor's size and speed figures also vary between documents. The model card gives 5.95 GB and a reduction of about 9.0x for the PTQ1_0 file, while the whitepaper gives 5.93 GB and about 9.1x against a 53.80 GB FP16 baseline.^[19][20] In the card's llama-bench measurements (batch size 1, 128 generated tokens), decode speed is 129.9 tokens per second on an RTX 5090 with PQ2_0 (120.5 with PTQ1_0) and 28.1 tokens per second on an Apple M5 Pro laptop.^[19] The press release gives up to 143 tokens per second on the RTX 5090, which matches the whitepaper's 142.5 tokens per second for PQ2_0 (134.4 for PTQ1_0, and 27.7 on the M5 Pro).^[8][20]

Memory beyond the weights

The 5.95 GB figure covers the language model's weights only. PrismML's own Bonsai 27B model card explains that a device must hold peak memory: the weights plus the KV cache, activations and runtime buffers (about 1.3 GB across backends). For 1-bit Bonsai 27B it measured 3.79 GB of weights, 5.2 GB peak at a 4K-token context and 11.6 GB at 100K tokens with an FP16 cache, and lower peaks with a 4-bit cache.^[23] The company's Bonsai 27B announcement makes the same point about phones: a 12 GB iPhone offers an app about 6 GB, which the model shares with its KV cache and activations.^[14]

Reception and evaluation

Coverage has mostly relayed PrismML's own numbers. CNBC reported that analysts cautioned the claims "still need to be proven outside controlled demonstrations". Tarun Pathak of Counterpoint Research said "the ultimate test will be millions of queries, thousands of device combinations and robust testing at scale", and Phil Solis of IDC said power consumption may be the biggest open question.^[5] The Register noted that Qwen3 8B scored somewhat ahead of 1-bit Bonsai 8B on the company's benchmark list.^[3] Hassibi told CNBC that the models typically lose a few percentage points of overall performance, with factual recall degrading before reasoning, math and coding.^[5]

At least one competitor has published a comparison. Underdog (Conway Research), which publishes its own open-weight 2-bit builds of Qwen3.8-27B, ran Ternary Bonsai 2 27B (PTQ1_0) through its tests at launch of its Underdog Saluki 27B. On Underdog's 120-task benchmark the Bonsai build passed 70 tasks against 84 for full-size Qwen3.8-27B and 88 for Saluki. On 100 parallel tool-calling tasks from the Berkeley Function Calling Leaderboard it passed 11 against 35 for the full-size model, and on a 50-issue SWE-bench Verified subset it fixed 29 against 33.^[36] These are a rival vendor's own measurements.

On Hugging Face, the Ternary-Bonsai-2-27B-gguf repository recorded 1,953,832 downloads in the 30 days to 30 October 2026 and 5,708,739 in total; the 42 repositories under the prism-ml account recorded about 24.6 million downloads in total.^[28]

Continue reading

• BitNetBitNet is a family of large language model architectures developed by Microsoft Research Asia that constrain the weights of a transformer to extremely low bit-widths: initially a single bit ({-1, +1}) and…Mentioned 3 times above
• QQuantizationQuantization is the representation of values from a large or continuous set by values from a smaller, usually finite set.Mentioned 3 times above
• Model CompressionModel compression is a family of techniques that reduce the size, memory footprint, and computational cost of machine learning models while preserving as much of their original accuracy as possible.Listed in See also
• OOn-device AIOn-device AI is the practice of running machine learning models on the phone, laptop, watch, or embedded board a person is actually using, instead of sending the input to a remote data center.Listed in See also
• EEdge AIEdge AI is the practice of running artificial intelligence models directly on the device that generates the data, such as a smartphone, laptop, camera, vehicle, or IoT sensorListed in See also
• QQwen3.8Qwen3.8 is the name Qwen uses for a model generation that includes hosted services and downloadable checkpoints.Mentioned 9 times above

More to read

• GGUF
• BitNet b1.58
• Underdog Saluki 27B
• llama.cpp
• Google
• Qwen3.6

Topics:AI CompaniesAI InferenceOpen Source AISmall Language Models

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7Prism ML. "PrismML Launches World's First 1-Bit AI Model to Redefine Intelligence at the Edge" (press release). Pasadena, 31 March 2026. prismml.com/...aunches-worlds-first-1-bit-ai-model (accessed 30 October 2026).
2. ^1 ^2 ^3Prism ML. "Announcing 1-bit Bonsai: The First Commercially Viable 1-bit LLMs" (blog). 31 March 2026. prismml.com/...bonsai-8b (accessed 30 October 2026).
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7Claburn, Thomas. "PrismML debuts energy-sipping 1-bit LLM in bid to free AI from the cloud." The Register, 4 April 2026. theregister.com/...prismml_1bit_llm
4. ^1 ^2 ^3Prism ML. "About" (team page) and careers listings. prismml.com/about (accessed 30 October 2026).
5. ^1 ^2 ^3 ^4 ^5 ^6Sigalos, MacKenzie. "Apple in talks with startup that shrinks AI models to run on an iPhone." CNBC, 14 July 2026. cnbc.com/...apple-prismml-ai-compression-iphone
6. ^1 ^2 ^3Bort, Julie. "PrismML hopes its tiny LLM will change how we all use AI." TechCrunch, 17 September 2026. techcrunch.com/...m-could-change-how-we-all-use-ai
7. ^1 ^2Bort, Julie. "PrismML brings its tiny LLMs to Qualcomm-powered smart glasses." TechCrunch, 24 September 2026. techcrunch.com/...o-qualcomm-powered-smart-glasses
8. ^1 ^2 ^3 ^4 ^5 ^6PrismML. "PrismML Launches Bonsai 2 27B, Its Most Capable Model Yet" (press release via PR Newswire). 17 September 2026. morningstar.com/...-27b-its-most-capable-model-yet
9. ^1 ^2 ^3Prism ML. "PrismML Introduces Ternary Bonsai Model Family" (press release). Pasadena, 16 April 2026. prismml.com/...roduces-ternary-bonsai-model-family (accessed 30 October 2026).
10. ^1 ^2 ^3Prism ML. "Introducing Ternary Bonsai: Top Intelligence at 1.58 Bits" (blog). April 2026. prismml.com/...ternary-bonsai (accessed 30 October 2026).
11. ^1 ^2Prism ML. "PrismML Releases Bonsai Image 4B" (press release). Pasadena, 26 May 2026. prismml.com/...prismml-releases-bonsai-image-4b (accessed 30 October 2026).
12. ^1 ^2 ^3Prism ML. "Introducing 1-bit and Ternary Bonsai Image 4B: Image Generation for Local Devices" (blog). May 2026. prismml.com/...bonsai-image-4b (accessed 30 October 2026).
13. ^1 ^2 ^3Prism ML. "PrismML Announces 1-bit Bonsai 27B: The First 27B Model to Run on a Phone" (press release). Pasadena, 14 July 2026. prismml.com/...prismml-releases-bonsai-27b (accessed 30 October 2026).
14. ^1 ^2 ^3 ^4Prism ML. "Announcing Bonsai 27B: The First 27B-Class Model to Run on a Phone" (blog). July 2026. prismml.com/...bonsai-27b (accessed 30 October 2026).
15. ^Prism ML. "Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint" (blog). September 2026. prismml.com/...bonsai-2-27b (accessed 30 October 2026).
16. ^1 ^2 ^3 ^4Prism ML. "PrismML Brings 1-Bit Bonsai Models to AI Smart Glasses Powered by Snapdragon" (press release). Pasadena, 23 September 2026. prismml.com/...smart-glasses-powered-by-snapdragon (accessed 30 October 2026).
17. ^1 ^2 ^3Prism ML. "PrismML Expands Bonsai 2 Family with 2B, 4B, and 9B Models for On-Device AI" (press release). Pasadena, 15 October 2026. prismml.com/...b-4b-and-9b-models-for-on-device-ai (accessed 30 October 2026).
18. ^Prism ML. "Introducing Ternary Bonsai 2 2B, 4B, and 9B: Bringing Local Agents to Everyday Devices" (blog). October 2026. prismml.com/...bonsai-2-family (accessed 30 October 2026).
19. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25prism-ml. "Bonsai 2 27B: GGUF" (model card, Ternary-Bonsai-2-27B-gguf). Hugging Face. huggingface.co/...Ternary-Bonsai-2-27B-gguf (accessed 30 October 2026).
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10PrismML. "Ternary Bonsai 2 27B: A 27B ternary reasoning model" (whitepaper). September 2026. github.com/...bonsai-2-27b-whitepaper.pdf (accessed 30 October 2026).
21. ^1 ^2 ^3 ^4 ^5prism-ml. "Bonsai-8B-GGUF-1bit" (model card). Hugging Face. huggingface.co/...Bonsai-8B-gguf (accessed 30 October 2026).
22. ^1 ^2 ^3prism-ml. "Ternary-Bonsai-8B-gguf" (model card); see also the Ternary-Bonsai-4B-gguf and Ternary-Bonsai-1.7B-gguf cards. Hugging Face. huggingface.co/...Ternary-Bonsai-8B-gguf (accessed 30 October 2026).
23. ^1 ^2 ^3 ^4prism-ml. "1-bit Bonsai 27B: GGUF" (model card, Bonsai-27B-gguf). Hugging Face. huggingface.co/...Bonsai-27B-gguf (accessed 30 October 2026).
24. ^prism-ml. "Ternary Bonsai 27B: GGUF" (model card). Hugging Face. huggingface.co/...Ternary-Bonsai-27B-gguf (accessed 30 October 2026).
25. ^1 ^2prism-ml. "Bonsai 2 Ternary 9B: GGUF" (model card); see also the 4B and 2B cards. Hugging Face. huggingface.co/...Ternary-Bonsai-2-9B-gguf (accessed 30 October 2026).
26. ^prism-ml. "Bonsai-4B-GGUF-1bit" (model card). Hugging Face. huggingface.co/...Bonsai-4B-gguf (accessed 30 October 2026).
27. ^prism-ml. "Bonsai-1.7B-GGUF-1bit" (model card). Hugging Face. huggingface.co/...Bonsai-1.7B-gguf (accessed 30 October 2026).
28. ^1 ^2Hugging Face. Model API listing for the prism-ml account (downloads, downloadsAllTime, createdAt). huggingface.co/...models (accessed 30 October 2026).
29. ^1 ^2GitHub. Repository records for PrismML-Eng/Bonsai-demo and PrismML-Eng/llama.cpp. github.com/...Bonsai-demo and github.com/...llama.cpp (accessed 30 October 2026).
30. ^1 ^2 ^3khosravipasha. "ggml: add Q1_0 1-bit quantization support (CPU)." llama.cpp pull request #21273, merged 6 April 2026. github.com/...21273
31. ^1 ^2 ^3khosravipasha. "ggml: add Q2_0 quantization support (CPU)." llama.cpp pull request #24448, merged 7 July 2026. github.com/...24448
32. ^Kemper, Jonathan. "Bonsai 27B is a full open reasoning model that fits on an iPhone." The Decoder, 15 July 2026. the-decoder.com/...ng-model-that-fits-on-an-iphone
33. ^Hall, Zac. "PrismML releases Bonsai 27B, claiming first major AI model of its size fit for iPhone." 9to5Mac, 14 July 2026. 9to5mac.com/...ai-model-of-its-size-fit-for-iphone
34. ^1 ^2 ^3"Apple is in talks with the startup shrinking a 27B AI model onto an iPhone." The Next Web, 15 July 2026. thenextweb.com/...-on-device-ai-compression-iphone
35. ^Dotson, Kyt. "PrismML launches Bonsai 2 27B, a high-intelligence AI model so small it fits on consumer hardware." SiliconANGLE, 18 September 2026. siliconangle.com/...l-it-fits-on-consumer-hardware
36. ^Underdog. "Meet Underdog Saluki 27B." 7 October 2026. underdog.ai/saluki (accessed 30 October 2026).
37. ^Hassibi, Babak; Stork, David G. "Second Order Derivatives for Network Pruning: Optimal Brain Surgeon." Advances in Neural Information Processing Systems 5 (NIPS 1992), pp. 164-171. CaltechAUTHORS record. resolver.caltech.edu/...AUTHORS:20150219-075206704
38. ^khosravipasha. "cuda: Q1_0 initial backend." llama.cpp pull request #21629, merged 15 April 2026. github.com/...21629
39. ^Ma, Shuming; Wang, Hongyu; et al. "The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits." arXiv:2402.17764, 27 February 2024. arxiv.org/...2402.17764

Show all 39 references