Code Llama 4 (70B & 400B) vs SAM 3 (Segment Anything Model 3)

“The primitive here is raw model weights you can actually run: no API wrapper, no rate limits, no vendor controlling your uptime. The DX bet Meta made is correct — drop weights on Hugging Face, let the ecosystem (vLLM, llama.cpp, Ollama) handle the serving layer. The moment of truth is spinning up a 70B quant locally or on a single A100, and that actually works without 12 env vars. The 400B is a different story — you're in multi-GPU territory fast — but the 70B is a genuine weekend-deployable primitive. The specific decision that earns the ship: function calling support baked in at the weight level means you're not duct-taping tool use on top after the fact.”

“The primitive is clean: promptable segmentation over images, video frames, and sparse 3D point clouds via a unified inference interface — no fine-tuning required. The DX bet Meta made is that developers want a composable foundation model they can drop into a pipeline, not a SaaS endpoint they have to negotiate with, and that bet is exactly right. Where SAM 1 required post-processing hacks to propagate masks across frames, SAM 3 handles temporal consistency natively, which eliminates a whole category of brittle glue code I've personally written. The specific technical decision that earns the ship: open weights with a documented Python API that doesn't require you to memorize a config file before you can run inference on a single image.”

“Direct competitors are GPT-4.1, Claude Sonnet 3.7, and Qwen2.5-Coder — all of which have closed weights or commercial restrictions. The specific scenario where Code Llama 4 breaks is enterprise fine-tuning at 400B scale: most teams can't afford the compute to actually adapt it, so they'll run 70B quantized and wonder why it doesn't hit benchmark numbers. The HumanEval and SWE-bench claims need scrutiny — Meta authored the eval setup, and 'state-of-the-art' on benchmarks designed around pass@1 on clean problems doesn't map cleanly to real codebases with legacy debt and ambiguous specs. What saves this from a skip: the permissive license is real, the Hugging Face availability is real, and the 70B model gives teams genuine pricing leverage against OpenAI. Prediction: this wins by being the baseline every fine-tune starts from, not by being the best raw model.”

“Direct competitors are SAM 2 (which this replaces), Grounded-SAM pipelines, and the growing cluster of closed segmentation APIs from Roboflow and Scale AI — SAM 3 beats all of them on cost (free) and beats most on video consistency without needing a separate tracker bolted on. The scenario where this breaks is 3D: 'preliminary point-cloud support' is doing a lot of work in that sentence, and anyone who tries to run this on dense LiDAR scans for autonomous driving will hit accuracy floors fast. What kills this in 12 months isn't a competitor — it's Meta's own next release; the model will be superseded, but the open-weights distribution model means SAM 3 stays useful in frozen production pipelines long after SAM 4 drops, which is the real moat here.”

“The thesis: by 2027, the majority of production code-generation inference runs on self-hosted open weights because closed API costs are structurally incompatible with the volume that agentic coding pipelines generate. Code Llama 4 is a direct bet on that trajectory, and the 70B/400B split is smart — it covers the 'runs on one node' use case and the 'we have a cluster' use case simultaneously. The second-order effect that matters most isn't cheaper completions — it's that fine-tuning on proprietary codebases becomes viable without shipping your IP to a third-party API. The trend line is the commoditization of inference hardware plus the normalization of multi-step coding agents; Code Llama 4 is on-time, not early. The future state where this is infrastructure: every mid-size engineering org runs a Code Llama 4 fine-tune on their own codebase as a first-class internal tool, same as they run their own CI.”

“The thesis SAM 3 bets on: by 2028, visual understanding is a commodity layer, and the developers who own application logic on top of open segmentation primitives will capture more value than those who depend on closed vision APIs. That's a plausible and falsifiable claim — it fails if frontier closed models (GPT-5V, Gemini Ultra vision) get cheap enough that the total cost of ownership for open weights (infra, latency tuning, versioning) exceeds the API bill. The second-order effect nobody is talking about: real-time video segmentation at this quality level unlocks sports analytics, retail foot-traffic analysis, and AR object persistence for teams that previously couldn't afford the compute or the licensing. SAM 3 is on-time to the open computer vision trend — not early, not late — and it's well-positioned because Meta's institutional commitment to open weights is a credible signal that this won't be quietly deprecated behind a paywall.”

“The buyer here isn't an individual — it's an engineering team with a cloud bill and a compliance department that doesn't want code leaving the perimeter. That's a real, funded budget: 'self-hosted AI' sits in infra, not experimental tooling. The moat question is where this gets complicated: Meta has no moat in the traditional sense, but the ecosystem lock-in comes from fine-tune artifacts and toolchain integrations that accumulate over time. The real business risk is that Meta releases Code Llama 5 in eight months and the 400B variant is immediately obsolete before most teams have even finished deploying it — the open-source cadence creates capability depreciation that's faster than enterprise adoption cycles. Still a ship because the pricing model — free weights, you pay for compute you'd be paying for anyway — is the only model that survives contact with a CFO asking why you're paying per-token for internal tooling.”

“The job-to-be-done is singular and clear: give me accurate object masks from a prompt, across video frames, without training a custom model. SAM 3 nails that job for images and mostly nails it for video; the 3D support is more 'tech preview' than 'shipped feature' and shouldn't factor into adoption decisions today. Onboarding is as fast as cloning a repo and running the example notebook — value in under 5 minutes if you have a GPU, which is the right bar for a developer-facing research artifact. The product opinion is strong: Meta has decided that promptable segmentation (clicks, boxes, text) is the right interaction model rather than category-specific fine-tuned heads, and every design decision flows from that commitment — which is exactly the kind of opinionated stance that makes a tool actually useful rather than infinitely configurable and practically useless.”