Overview

VIDRAFT is a 12-person startup operating out of Seoul AI Hub, a Seoul Metropolitan Government facility, founded in 2024 by CEO Min-sik Kim, who was a fourth-year computer engineering student at Korea Cyber University before taking leave to run the company. It has no large funding round to point to. What it has instead is a NIPA GPU allocation rated "Excellent" for 2026, KOITA recognition as a corporate research institute, and access to 24 GPUs (16 Blackwell B200s and 8 H200s), which is a modest fleet by frontier-lab standards but enough to run its core product: Darwin, a gradient-free model-merging engine that recombines the weights of existing checkpoints instead of training new ones.

Darwin-36B-Opus, the model POCKET-35B is quantized from, is the highest-scoring member of that family. VIDRAFT bred it from two parents that share a common base: Qwen/Qwen3.6-35B-A3B supplied the MoE architecture and expert topology, while hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled, a LoRA fine-tune trained on 14,233 chain-of-thought traces distilled from Claude Opus 4.6, supplied the reasoning behavior. Darwin V7 merges the two through what VIDRAFT calls per-tensor compatibility analysis rather than any further gradient descent, and the whole process reportedly finishes in under 10 minutes on a single B200 GPU.

POCKET-35B strips that merged model down for hardware that has none of that. It ships as four GGUF quantizations, from a 21.2GB top-quality build to an 8.24GB build small enough for a 16GB RAM box, all built to load in unmodified llama.cpp rather than a custom fork.

POCKET-35B's headline CPU numbers come from a 16-thread Xeon server, the kind of hardware VIDRAFT is targeting instead of a GPU. Source: commons.wikimedia.org

Key Specifications

The parameter count is worth a note. VIDRAFT markets this as a "35B" model, and the GGUF file metadata reports 34,660,610,688 total parameters, close enough to round to 35B and consistent with the Darwin-36B-Opus parent's own spec sheet. The architecture field in that same metadata is labeled qwen35moe, a naming holdover from the underlying model class rather than a claim about which Qwen generation it descends from. Darwin-36B-Opus's own model card is explicit that the base parent is Qwen 3.6-35B-A3B, not 3.5, even though POCKET's own README describes the lineage as "Qwen3.5-family MoE" in one place. That's a real inconsistency between VIDRAFT's own documents, and it's the kind of detail worth checking yourself before citing a model's lineage secondhand.

Benchmark Performance

VIDRAFT benchmarks POCKET-35B's IQ1_M build against Bonsai-27B Q1_0, the 1-bit build of PrismML's Bonsai 27B, on the same hardware with the same stock llama.cpp binary. The comparison is self-reported by VIDRAFT, not independently reproduced, but the methodology (same tools, same machine, disclosed losses) is more transparent than most vendor benchmarks.

"MoE prefill wakes every expert, so sparsity stops helping there," VIDRAFT writes in its own README, describing why POCKET loses the GPU prompt-processing test.

That line matters more than the wins do. Sparse MoE only pays off during generation, where the router activates a handful of the 256 experts per token. Prompt processing has to touch a much wider slice of the expert pool across a long input, so the "3B active parameters" pitch collapses on an H100 doing prefill, and Bonsai's dense-leaning 27B design comes out over twice as fast there. On a laptop, that gap disappears. VIDRAFT's own numbers show POCKET winning every axis on a MacBook M3 Pro, including prompt processing (240.7 versus 73.4 tok/s over Metal), because there's no longer enough raw GPU throughput for a dense model to out-brute-force a sparse one.

Reasoning quality tells a separate story. Quantizing the model costs real accuracy on GPQA Diamond's 198 questions under greedy decoding: the unquantized Qwen 3.6-35B-A3B base scores 73.2%, POCKET-35B's Q4_K_M build drops to 68.7%, and Q2_K falls further to 60.1%. VIDRAFT hasn't published a GPQA figure for the smallest IQ1_M build. That 73.2% baseline is also worth cross-checking against our scientific reasoning leaderboard: Alibaba's own materials for Qwen 3.6-35B-A3B report 86.0% GPQA elsewhere, a 13-point gap from VIDRAFT's greedy run on the same base model that has nothing to do with quantization and everything to do with eval harness and prompting differences. Treat any single GPQA number, from any vendor, as harness-dependent until proven otherwise.

Key Capabilities

Where the parent model's score actually comes from

Darwin-36B-Opus's headline 88.4% GPQA Diamond score, the number that put VIDRAFT on Hugging Face's community leaderboard with 397B-parameter dense models, isn't a single-pass result. VIDRAFT's own evaluation methodology scores it in two phases: a greedy first pass across all 198 questions lands at 73.2% (145 correct), and a second pass re-runs every miss with 8-way majority voting plus a 16-vote tiebreaker round for close calls, lifting the final score to 88.4% (175 correct). That's a 15.2 percentage point gain from stochastic retry alone. POCKET-35B doesn't inherit that benefit, since nobody runs an 8-way majority vote on a phone; its GPQA numbers above are single-pass greedy, which is the fairer comparison for how the model actually gets used on-device.

Positioned against a real leaderboard, not just Bonsai

VIDRAFT's own README places Darwin-36B-Opus tied for third on an April 23, 2026 snapshot of Hugging Face's GPQA Diamond leaderboard, level with the 397B-parameter Qwen3.5-397B-A17B and behind two entries from TNSA. A BigGo Finance report published five days later, on April 28, put the same model in fifth place. Leaderboard rank shifts fast as new entries land, so treat any specific ordinal as a snapshot rather than a fixed fact. The score itself (88.4%, Pass 2) has held up across VIDRAFT's own model card, its Hugging Face metadata, and third-party coverage.

The underlying research is documented in arXiv:2605.14386, "Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning," published May 14, 2026 by a seven-person VIDRAFT-affiliated author group as a NeurIPS 2026 submission. Its headline claim is actually about a different sibling: Darwin-27B-Opus at 86.9% GPQA Diamond, ranked 6th among 1,252 assessed models. Darwin-36B-Opus, the model POCKET-35B descends from, is the newer and higher-scoring release that came after the paper's own reported results.

Every device tier gets a different repo, not just a different quant

POCKET's phone-capable builds live in sibling repos, not in POCKET-35B-GGUF itself - the iPhone MLX build is a separate, smaller download. Source: unsplash.com

It's worth being precise about what "POCKET" actually ships, because the collection is bigger than the one repo this article covers. POCKET-35B-GGUF, the subject here, targets PC and server CPUs with no GPU. A sibling repo, POCKET-KR-MLX, ships a 2-bit Apple MLX build sized for iPhone and Mac and tuned for Korean. POCKET-EN-GGUF covers English on both PC and phone with a 5.3GB "iPhone-mix" build meant to run through the PocketPal app. There's also POCKET-26B, a Gemma4-based sibling scoring 67% on GPQA Diamond that trades some quality for broader app compatibility today. VIDRAFT's iPhone and Android throughput claims apply to those sibling builds, not to POCKET-35B-GGUF's own four files, which VIDRAFT itself lists as not yet benchmarked on phone hardware.

Pricing and Availability

Every POCKET-35B tier is sized against system RAM, not VRAM - the whole pitch is that no GPU line item is needed at all. Source: unsplash.com

POCKET-35B-GGUF is free, hosted on Hugging Face under FINAL-Bench, released under Apache 2.0 with no usage restrictions and no API tier. There's nothing to subscribe to. The only cost is hardware, and VIDRAFT is explicit about matching each quantization to a RAM budget rather than a GPU tier:

• Q4_K_M (21.2GB) - top quality, needs roughly 32GB RAM
• Q3_K_M (16.8GB) - a 24GB RAM machine
• Q2_K (12.9GB, VIDRAFT's recommended "daily driver") - 16 to 24GB RAM
• IQ1_M (8.24GB) - the smallest full build, fits a 16GB RAM box

Quickstart is a single llama-cli invocation with -ngl 0 to force CPU-only execution, no custom build flags:

llama-cli -m POCKET-35B-Q2_K.gguf -p "your prompt" -ngl 0 -t 8

VIDRAFT recommends capping -t at physical core count rather than logical thread count, since oversubscribing threads reportedly slows generation rather than speeding it up. For anyone weighing this against a conventional quantized deployment, our guide to running open-source LLMs locally and the LLM quantization impact leaderboard both cover the tradeoffs of running Q2-class quantizations more broadly.

Strengths and Weaknesses

Strengths

• Runs on stock llama.cpp, Ollama, and LM Studio with no fork or custom build, unlike VIDRAFT's own claim that a comparable Bonsai GGUF build fails to load without a PrismML-patched runtime
• CPU generation beats the comparable Bonsai-27B build by 2.69x on a 16-thread Xeon and by more than 3x on a MacBook M3 Pro's CPU cores
• Retains the full 262,144-token context window inherited from Qwen 3.6-35B-A3B, confirmed directly in the GGUF file metadata rather than just marketing copy
• Four quantization tiers give a real choice between quality (Q4_K_M) and footprint (IQ1_M) rather than one-size-fits-all compression
• Apache 2.0 licensing with no usage restrictions and no pricing tier to negotiate

Weaknesses

• GPQA Diamond drops from the base model's 73.2% to 68.7% at Q4_K_M and 60.1% at Q2_K; quality loss is real and gets worse fast at lower bit-widths
• Loses to Bonsai-27B on GPU prompt processing by more than half (753 versus 1,816 tok/s on an H100), since sparse MoE routing doesn't help during prefill
• VIDRAFT's own documentation contains a lineage inconsistency, describing the base architecture as both "Qwen3.5-family" and explicitly Qwen 3.6-35B-A3B in different documents
• The 88.4% flagship GPQA score belongs to the unquantized Darwin-36B-Opus parent under expensive 8-way majority voting, not to anything POCKET-35B can reproduce in a single pass on a phone
• No GPQA figure disclosed for the smallest IQ1_M quantization, the build most likely to run on the lowest-memory devices

Related Coverage

• Qwen 3.6-35B-A3B model card - the architecture and base weights Darwin-36B-Opus was bred from
• Bonsai 27B model card - the PrismML compression model VIDRAFT benchmarks POCKET-35B against directly
• Edge and Mobile LLM Leaderboard - on-device throughput rankings across phones, laptops, and GPU-free hardware
• LLM Quantization Impact Leaderboard - how much quality other models lose across INT4 through 1-bit formats
• Scientific Reasoning LLM Leaderboard - GPQA Diamond rankings for full context on where 73.2% and 88.4% sit
• How to Run Open-Source LLMs Locally - setup guidance for llama.cpp, Ollama, and MLX deployments

Sources

• FINAL-Bench/POCKET-35B-GGUF on Hugging Face
• FINAL-Bench/Darwin-36B-Opus on Hugging Face
• SeaWolf-AI - POCKET launch post on Hugging Face
• Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning - arXiv:2605.14386
• Hugging Face GPQA Leaderboard Dominated by China and South Korea Models; VIDRAFT's Darwin Surges - BigGo Finance
• VIDRAFT company site
• Qwen/Qwen3.6-35B-A3B on Hugging Face
• hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled on Hugging Face

The twist is how VIDRAFT got there. POCKET wasn't trained. It was bred.

The Claims, Side by Side

A Model Bred, Not Trained

VIDRAFT's core technology isn't POCKET, it's the engine that produced its parent. Called Darwin V7, the system treats existing open model checkpoints as a genetic pool and recombines their weight tensors without a single gradient step. Per-tensor compatibility analysis decides which components can merge safely, an automated recombination pass builds the offspring, and a multi-phase benchmark run verifies the result. VIDRAFT says the whole breeding cycle takes under 10 minutes on one GPU, a fraction of what pretraining a model this size normally costs.

Darwin-36B-Opus, the parent behind POCKET, came from crossing Qwen3.6-35B-A3B, Alibaba's hybrid Gated DeltaNet/attention MoE base, with a second model fine-tuned on Claude Opus 4.6 reasoning traces. The recombination logic biases toward the reasoning parent for components tied to chain-of-thought and toward the base parent everywhere else. The method, detailed in an arXiv paper titled "Darwin Family: MRI-Trust-Weighted Evolutionary Merging," calls this a 14-dimensional "adaptive merge genome," and it's now sitting in front of NeurIPS reviewers.

The Leaderboard Climb

The technique's real proof point predates POCKET by three months. On Hugging Face's GPQA Diamond leaderboard, a 1,252-model ranking of PhD-level science reasoning, VIDRAFT's Darwin family occupies every one of South Korea's five entries in the top 21, with Darwin-28B-Opus taking 3rd place at 88.89%, ahead of a dense 397-billion-parameter competitor roughly ten times its size. China holds the other 16 spots in that top tier.

VIDRAFT operates out of Seoul AI Hub, a city-backed startup facility, and became the only South Korean startup to get a direct collaboration proposal from Hugging Face in March 2026. Source: unsplash.com

"Given the cost and resources required to build large models from scratch, we needed a different approach. Combining models turned out to deliver real performance gains," VIDRAFT CEO Kim Min-sik said of the strategy.

VIDRAFT was founded in 2024, is backed by South Korea's National IT Industry Promotion Agency, and has spun its Darwin lineage into roughly 380 derivative models on the open-source leaderboard ecosystem. For a team this size to out-rank labs with dedicated pretraining budgets, on a scientific reasoning benchmark no less, is the kind of underdog result that gets attention. Whether it holds up depends entirely on how you read the number.

What They Measured, and What They Didn't

The Multi-Pass Gap

The 88.4% headline score isn't what Darwin-36B-Opus gets on the first try. VIDRAFT's own model card shows Pass 1, greedy decoding, lands at 73.2%. Pass 2, which retries with temperature-based sampling and takes the majority vote across attempts, recovers another 15.2 points. That's a legitimate technique, plenty of labs report best-of-N numbers, but it means the leaderboard-topping figure describes a different, more expensive inference process than a single query. And the greedy number, the one closest to how most people actually run a chatbot, is identical to the unmerged Qwen3.6-35B-A3B base model. The breeding process didn't move that needle at all.

The Compression Cost

POCKET inherits the same ceiling and loses more on the way down. Its Q4_K_M build, the largest at 21.2GB, scores 68.7% greedy. The 12.9GB Q2_K variant VIDRAFT recommends as the "best value" option drops to 60.1%. When a Hugging Face commenter, Dipankar Sarkar, asked why POCKET loses to a dense competitor on H100 prompt processing (0.41x) but wins on the same operation on a MacBook M3 Pro, the mismatch points to something POCKET's own documentation admits: MoE sparsity helps token generation, where only 8 of 256 experts activate, but prompt processing wakes every expert regardless, erasing the advantage on hardware built for dense matrix throughput.

None of this means the numbers are fabricated. VIDRAFT publishes its quantization variants, its perplexity scores per language, and its benchmark harness openly, and the whole stack runs on stock llama.cpp anyone can audit. It means the marketing framing (top-5 on a leaderboard of 1,252 models) and the practical framing (what a single query on your phone actually returns) are two different claims wearing the same benchmark number.

VIDRAFT frames Darwin's weight-recombination process as evolutionary "breeding" rather than training - a framing this piece borrows because it's technically apt, not because it's marketing copy. Source: unsplash.com

Should You Care?

If the goal is a truly capable reasoning model that fits on a phone or a GPU-less mini PC, POCKET-35B is a real option, not vaporware. Apache 2.0 licensing, honest per-quant benchmarks, and 2-3x CPU throughput over a comparably sized Bonsai-27B are useful wins for anyone doing on-device inference. Just don't treat the GPQA Diamond rank as a description of what you'll get from a single prompt on a 16GB laptop. Read it the way you'd read any contested leaderboard result: as the best case under generous sampling, not the expected one.

For VIDRAFT, the more durable story isn't the leaderboard slot, it's the method. Breeding a 35B model from existing checkpoints in under 10 minutes on a single GPU, then compressing the result for phones, is a genuinely different way to compete with labs spending hundreds of millions on pretraining. Whether Darwin V7 keeps producing gains as the genetic pool of available open models gets picked over is the open question worth watching, not this month's leaderboard screenshot.

Sources:

• POCKET-35B-GGUF model card
• Darwin-36B-Opus model card
• SeaWolf-AI POCKET announcement post
• Darwin Family: MRI-Trust-Weighted Evolutionary Merging (arXiv)
• BigGo Finance: Hugging Face GPQA Leaderboard Dominated by China and South Korea Models
• KMJ: China and South Korea Dominate Hugging Face GPQA Leaderboard