A model sized for the machine on your desk

Recursive positions Flint for suitably configured 48GB-class hardware: recent Apple Silicon Macs and 48GB NVIDIA GPUs. The figure to handle carefully is memory. The 4-bit weights take about 23GB, but that is only part of the requirement — runtime overhead, the KV cache and long contexts all add to the total, and a long session needs more than the download size suggests. The full BF16 checkpoint is a different class of deployment: the model card puts it at about 80GB, served on one B300 or H200 GPU per replica with Recursive's SGLang kit.

The speed claims need the same care. Recursive reports more than 125 tokens per second on an M4 Pro at short contexts, and more than 150 on a 64GB NVIDIA RTX Spark, using its own optimised inference stack. Those are company-reported figures for a specific hardware and runtime combination at short contexts — not a guaranteed speed across hardware, runtimes or session lengths, and throughput falls as a session grows. Flint works with the local tools that already exist for this — llama.cpp, Ollama, Unsloth, LM Studio, vLLM and SGLang — and Recursive is also releasing its own engine, rec, tuned for the M4 Pro and M5 Pro and Max chips.

What local deployment does change is where the code goes. Run entirely locally, without network-connected tools or services, Flint can process private code without sending prompts to a hosted provider, and there is no per-token bill. That is a property of the deployment, not just of the model: developers still need to inspect the privacy and execution behaviour of the agent runtime they choose, because external tools, web retrieval, telemetry or hosted integrations can still send information off the device.

The model card carries the matching caution. Flint's fine-tuning included safety-oriented examples, but the card states that the model can still produce incorrect or harmful content, faulty code or inappropriate tool calls, and that applications using it should enforce their own permissions and execution boundaries — structured tool-call support does not prevent the model from choosing an unsafe or incorrect action. Generated code needs testing and review before it is applied.

The context figures need a careful read. The launch post advertises windows of up to 512k tokens, with a recommendation to stay within 256k. The model card documents 262,144 tokens, shared between prompt and generated output. Until Recursive reconciles the two, plan around the card: about 256k of usable context, still large for a local coding model, but half the headline number.

The scores, sorted by who measured them

Recursive's central claim is precise, and its precision is the story. Flint, the company says, leads mixture-of-experts models with fewer than 50 billion total parameters on agentic coding benchmarks — a mean across six tests, in a category with the boundaries drawn around the product: MoE only, under 50B only. The claim is company-defined and attributed to Recursive. Within those boundaries the reported numbers are strong: Recursive also reports 80.3% on LiveCodeBench v6 and 95.1% on IFEval, all from its own release evaluations.

The more useful reference point is Recursive's reported 74.2% on SWE-bench Verified. Verified is a recognised benchmark with results published by other developers, which makes it the better anchor — but it is still a company-run evaluation, and harness, sampling and scoring differences can affect comparisons. The model card states its conditions: temperature 1.0, top-p 0.95 and top-k 20, with agentic scores averaged over at least five attempts, and it warns that evaluation setups vary across sources. Treat the 74.2% as Recursive's measurement on a public benchmark, not as an independently validated result.

Then there is the 89.7%. That figure is for SWE-Bench Pro v2, a variant whose competitor scores Recursive measured itself, and it must not be compared directly with public SWE-bench Pro leaderboards: a reader who sets it next to published Pro results from other labs is comparing two different measurements. The 89.7% is the number requiring the strongest qualification — not an invalid score, but the one carrying the most conditions.

Recursive's own chart supplies the counterweight. On Terminal-Bench 2.1, which tests an agent working in a real terminal, the chart shows Flint at 54.9% against dense Qwen 3.8 at 73.0% — a gap of 18.1 percentage points. Flint's sparse architecture helps reduce the computation per token, but the benchmark does not establish how much of that gap comes from the architecture: training data, post-training and agent configuration can all contribute. The comparison is Recursive's own chart, not an independent head-to-head. Recursive's guidance fits the same picture: use Flint for daily local coding, break complex work into smaller subtasks, give explicit instructions, and expect to guide it more than the large closed models.

The one independent hands-on account so far is encouraging and small. Wired's Will Knight ran Flint on a 48GB Mac, built a to-do app and a web-research tool, and reported that it handled every request, where smaller local models he had tried were slower or error-prone. That account reaches this article secondhand, via TechCrunch's reporting of it: one reporter's afternoon, not a benchmark. It is still more than a launch chart.

The builder is the bigger claim

Flint is Recursive's first model, and the company is using it to make a claim about process. Its Eureka Machine proposes ideas, implements them, tests them and decides what to try next, running directions in parallel. Recursive says the system contributed across the stack, and three of its contributions are specific enough to check against the artefact itself.

On architecture, the system started from a baseline of 35 billion total parameters with 3 billion active, searched expert configurations, and settled on 40 billion total with 2.5 billion active. On quantization, it did something against common practice: it quantized the output layer to 4 bits, then recovered accuracy with an offline self-distillation stage in which the quantized model trained on its own pre-sampled responses, with the 16-bit model as teacher. The resulting NVFP4 build scores within 0.7 points of the 16-bit model across Recursive's benchmarks. On kernels, it wrote hardware-specific code, including a Metal megakernel on Apple chips that fuses the entire decode pass.

What Recursive has not provided is the number that would make the automation claim testable: how much development time the Eureka Machine saved. TechCrunch notes the company did not say how much time Eureka saved its researchers. Under four months from scratch, by a small team, is fast for a competitive open model — but the time saved is unquantified, and whether the machine did the work, accelerated the work, or mostly searched well inside a human-set plan is not yet answerable from what has been published.

That distinction matters for what to watch next. Recursive's stated bet is compounding: each model makes the Eureka Machine a better researcher, which makes the next model faster to build. The evidence that separates that reading from a well-run small-team launch is the second model — and elapsed time alone is not the evidence. What would count is reproducible contributions, the human engineering effort involved, the compute resources used and a task of comparable difficulty, alongside the calendar. For now, what a reader can verify is narrower and still useful: a downloadable coding model under a permissive licence, company-reported short-context speeds above 125 tokens per second on an M4 Pro, and benchmark scores that reward reading the footnotes before repeating the headline.

Sources

  1. [1] Recursive, "Introducing Flint", launch post, 21 October 2026 — read directly for this article: 40B total / 2.5B active, under four months, the sub-50B MoE category claim as a mean across six tests, speed claims at short contexts, the up-to-512k context claim with the within-256k recommendation, Eureka Machine contributions to architecture, quantization and kernels, Hugging Face release under Apache 2.0, supported frameworks and the rec engineRead source
  2. [2] RecursiveSI/flint model card, Hugging Face — read directly for this article: 39.8B total parameters, about 2.5B active including the output head and about 2.2B active body, 49 layers, 256 routed experts with 6 per token plus a shared expert, documented context of 262,144 tokens shared between prompt and output, the about-80GB BF16 checkpoint with B300/H200 serving guidance, Apache 2.0 licence, the stated scores with evaluation conditions (temperature 1.0, top-p 0.95, top-k 20, agentic scores averaged over at least five attempts, setups vary across sources) and the safety note on faulty code and unsafe tool callsRead source
  3. [3] TechCrunch, "Recursive launches Flint, a laptop coding model it says was largely built by AI", published 21 October 2026, updated 22 October 2026 — company background ($650M at a $4.65B valuation from GV, NVIDIA and AMD Ventures, CEO Richard Socher), Recursive's about-131 tokens/sec M4 Pro test figure against 58–69 for comparable models, Wired's hands-on account as reported by TechCrunch, and Recursive not saying how much time Eureka savedRead source
  4. [4] MarkTechPost, "Recursive Releases Flint, an Open-Weight 40B MoE Local Coding Model for Agents with 2.5B Active Params", 21 October 2026 — launch-day coverage of the model card, including the note that SWE-Bench Pro v2 scores are Recursive-measured and not comparable to the public SWE-bench ProRead source