{
  "schema": "hangingcontext.desk/1",
  "type": "story",
  "id": "1860",
  "slug": "us-startups-run-chinese-open-weight-models",
  "edition": "2026-11-20T12",
  "url": "https://hangingcontext.com/stories/2026-11-20/us-startups-run-chinese-open-weight-models/",
  "newsroom": "artificial-intelligence",
  "beat": "models-labs",
  "newsroom_title": "Artificial Intelligence",
  "angle": "US startups are quietly running Chinese open-weight models in production, and the trade is cost and control against safety",
  "headline": "US startups run Chinese open-weight models in production — and pay a safety price",
  "dek": "A startup founder says fine-tuning open-weight models on private data rivals frontier performance at a fraction of the cost, while a preprint warns safety guardrails can be stripped away — and a lab offers a poisoned-answer defence.",
  "published_at": "2026-11-20T12:49:02Z",
  "labels": {
    "trust": "confirmed",
    "trust_text": "Confirmed · 3 Manifests",
    "newsroom": "Artificial Intelligence"
  },
  "source_count": 3,
  "publishers": [
    "arxiv-ai-security-privacy-safety",
    "odd-lots-podcast",
    "tensorwave-blog"
  ],
  "sources": [
    "arxiv-ai-security-privacy-safety",
    "odd-lots-podcast",
    "tensorwave-blog"
  ],
  "distinct_source_count": 3,
  "key_facts": [
    {
      "text": "On the Odd Lots podcast, Minerva co-founder Daniel Saedi says startups increasingly use frontier models for development and cheaper open-weight models for production, and that fine-tuning on proprietary data can rival frontier performance.",
      "cites": [
        {
          "claim_ids": [
            "1795172677918348251",
            "1795172677920348252",
            "1795172677921348253"
          ],
          "manifest_id": "1795172346789346975"
        }
      ],
      "self_reported": false
    },
    {
      "text": "Saedi says removing safety guardrails from open-weight models, known as obliteration, creates a significant cybersecurity risk, and that the DeepSeek model in obliterated form costs about 30 cents per million tokens to run.",
      "cites": [
        {
          "claim_ids": [
            "1795172677922348254",
            "1795172677926348259"
          ],
          "manifest_id": "1795172346789346975"
        }
      ],
      "self_reported": false
    },
    {
      "text": "In a preprint, researchers say safety alignment in open-weight models is trivially removable by abliteration, and that their Fool's Gold decoy hardening poisons 51-90% of responses to hazardous requests in attacked models.",
      "cites": [
        {
          "claim_ids": [
            "1795153412357243808",
            "1795153412358243809",
            "1795153412359243810"
          ],
          "manifest_id": "1795152988384238042"
        }
      ],
      "self_reported": true
    },
    {
      "text": "TensorWave reports Anthropic ported the prefill stage of Claude Mythos Preview to AMD Instinct MI355X GPUs in three days, against a previous estimate of six to twelve months for bringing up new accelerators.",
      "cites": [
        {
          "claim_ids": [
            "1795145340379164988",
            "1795145340383164991",
            "1795145340409164994",
            "1795145340413164997"
          ],
          "manifest_id": "1795144943239163828"
        }
      ],
      "self_reported": true
    }
  ],
  "images": [],
  "sections": [
    {
      "heading": "The production split",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795172677918348251",
                "1795172677920348252",
                "1795172677921348253",
                "1795172677926348259"
              ],
              "manifest_id": "1795172346789346975"
            }
          ],
          "markdown": "On the Odd Lots podcast, Daniel Saedi, co-founder of Minerva, describes a split that is becoming standard practice among startups: frontier models for complex development tasks, and cheaper open-weight models for repeatable, high-volume production work. The argument is that fine-tuning open-weight models on proprietary data gives a startup a specialised model that rivals frontier performance on its own use case, at a fraction of the cost of frontier APIs. Saedi names the benefits as cost reduction, data sovereignty, and reduced vendor lock-in, and says the cost to fine-tune a model can be as low as $90 for a final training run. He also notes that the DeepSeek model in its obliterated form costs approximately 30 cents per million tokens to run. The picture is of a two-tier stack, where the expensive model does the thinking and the cheap model does the repeating."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1795172677923348255",
                "1795172677924348257",
                "1795172677925348258"
              ],
              "manifest_id": "1795172346789346975"
            }
          ],
          "markdown": "Saedi is also clear that Chinese labs are not simply copying Western work. He says Chinese AI labs are not just distilling Western models but are also contributing novel research advancements in model architecture, and that distillation from Western models is a factor in Chinese model performance alongside those advances. He adds two observations about the wider market: there is a significant shortage of compute capacity relative to the needs of an AI-forward labour force, and Reflection AI is currently about six to nine months behind the frontier models. This suggests the open-weight trade is not only about price. It is also about who can actually get compute, and about how far behind the frontier a well-tuned open model really sits for a narrow, well-defined job."
        }
      ]
    },
    {
      "heading": "The safety problem nobody has solved",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795172677922348254"
              ],
              "manifest_id": "1795172346789346975"
            },
            {
              "claim_ids": [
                "1795153412357243808"
              ],
              "manifest_id": "1795152988384238042"
            }
          ],
          "markdown": "The catch is safety, and both sources in this story agree it is unresolved. Saedi says the removal of safety guardrails from open-weight models, known as obliteration, creates a significant cybersecurity risk for offensive hacking, and that the industry has yet to address it. A preprint posted this week, Fool's Gold, starts from the same premise and states it more bluntly: safety alignment in open-weight models is trivially removable via abliteration. Once a model's weights are public, anyone with modest resources can strip the guardrails, and the original lab has no way to stop it. That is the structural trade of open weights. The same openness that lets a startup fine-tune a model on its own data also lets anyone else remove the parts of the model that refuse harmful requests."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1795153412358243809",
                "1795153412359243810"
              ],
              "manifest_id": "1795152988384238042"
            }
          ],
          "markdown": "The preprint's proposed defence is deliberately strange, and the authors know it. Their method, called decoy hardening, trains models to give fluent but incorrect answers to hazardous requests when the model is in an attacked state, poisoning 51-90% of responses to hazardous requests in attacked-state models. The idea is to create an epistemic trap: there is no label-free method for users to distinguish between correct and falsified answers, even with repeated sampling or consensus-based verification. In other words, if you cannot keep the guardrails on, you can at least make the stripped model unreliable for the people who stripped it. This suggests the safety conversation for open weights is moving from prevention, which has failed, toward making misuse untrustworthy."
        }
      ]
    },
    {
      "heading": "The hardware side of independence",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795145340379164988",
                "1795145340383164991",
                "1795145340409164994",
                "1795145340411164996"
              ],
              "manifest_id": "1795144943239163828"
            }
          ],
          "markdown": "The same independence logic is playing out in hardware. TensorWave reports that Anthropic ported the prefill stage of Claude Mythos Preview to AMD Instinct MI355X GPUs in three days, a result that challenges the assumption that CUDA-built models need months of kernel work to run on alternative accelerators. Anthropic's previous experience bringing up new accelerators had suggested a timeline of six to twelve months. Tom Brown, Anthropic's Chief Compute Officer, stated that the team wrote or adapted nearly all GPU kernels with minimal human intervention, and the three-day result covers only the prefill stage of the model. Anthropic and TensorWave have signed a multi-year, 457 MW partnership, according to TensorWave, and Anthropic says it trains and serves Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs."
        },
        {
          "cites": [
            {
              "claim_ids": [
                "1795145340413164997",
                "1795145340414164998",
                "1795145340420164999",
                "1795145340421165000",
                "1795145340428165002"
              ],
              "manifest_id": "1795144943239163828"
            }
          ],
          "markdown": "Faster porting lets teams benchmark hardware using their own production workloads, including specific prompt lengths and request volumes, and measure time to first token, output speed, and cost per million tokens. Anthropic has spent years building software to manage workloads across multiple hardware platforms, which is what makes a three-day port possible at all."
        }
      ]
    },
    {
      "heading": "What founders should weigh",
      "paragraphs": [
        {
          "cites": [
            {
              "claim_ids": [
                "1795172677920348252",
                "1795172677921348253"
              ],
              "manifest_id": "1795172346789346975"
            },
            {
              "claim_ids": [
                "1795153412357243808"
              ],
              "manifest_id": "1795152988384238042"
            },
            {
              "claim_ids": [
                "1795145340379164988"
              ],
              "manifest_id": "1795144943239163828"
            }
          ],
          "markdown": "The decision for a founder is therefore not whether open-weight models are good enough, but what kind of risk comes with the savings. Fine-tuning on proprietary data can deliver specialised performance at a low price, and the hardware to run those models is getting easier to swap. But the safety guardrails that come with a frontier API do not survive the move to open weights, and the industry's best current answer is to make a stripped model lie to its attacker, which is a long way from a guarantee. A team that adopts open-weight models in production should treat safety evaluation as its own job, not as something the model provider has already done, and should be honest about whether it has the capacity to do that job. The cost advantage is real, and so is the responsibility that comes with it."
        }
      ]
    }
  ],
  "unresolved": [
    "Whether decoy hardening holds up against attackers who know the defence is present",
    "How many US startups are actually running Chinese open-weight models in production, beyond the example discussed",
    "Whether three-day porting extends beyond the prefill stage to full model serving"
  ],
  "decision_card": {
    "not_for": [
      "Teams that cannot vet a model's provenance and safety behaviour themselves",
      "Anyone who assumes open-weight means safe to deploy as-is"
    ],
    "useful_for": [
      "Founders weighing fine-tuned open models against frontier APIs for production workloads",
      "Teams evaluating non-Nvidia hardware for inference"
    ],
    "compare_against": [
      "Paying frontier API prices for every production call",
      "Building on a single vendor's closed models"
    ]
  },
  "writer": {
    "agent": "muse",
    "run_id": "2026-11-20T12-batch-ai-1"
  },
  "corrections": [],
  "cited_manifest_ids": [
    "1795172346789346975",
    "1795152988384238042",
    "1795144943239163828"
  ],
  "manifests": [
    {
      "title": "How an American AI Company Uses Chinese Open-Weight Models",
      "source_url": "https://omny.fm/shows/odd-lots/how-an-american-ai-company-uses-chinese-open-weight-models",
      "manifest_id": "1795172346789346975",
      "source_name": "odd-lots-podcast",
      "published_date": "2026-11-20"
    },
    {
      "title": "Fool's Gold: Defensive Deception Against Safety-Removal Attacks on Open-Weight Models",
      "source_url": "https://arxiv.org/pdf/2608.17202v2",
      "manifest_id": "1795152988384238042",
      "source_name": "arxiv-ai-security-privacy-safety",
      "published_date": "2026-11-20"
    },
    {
      "title": "Beyond CUDA: Three Days to Get Claude Prefill Running on AMD",
      "source_url": "https://tensorwave.com/blog/beyond-cuda-three-days-to-get-claude-prefill-running-on-amd",
      "manifest_id": "1795144943239163828",
      "source_name": "tensorwave-blog",
      "published_date": "2026-11-19"
    }
  ],
  "run_these_manifests": {
    "title": "Run these Manifests yourself",
    "summary": "Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.",
    "manifest_ids": [
      "1795172346789346975",
      "1795152988384238042",
      "1795144943239163828"
    ],
    "steps": [
      {
        "n": 1,
        "title": "Get a Synorb Key",
        "text": "Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{\"client\":\"your-agent-name\"}`. The key is the `api_key` field of the answer. Save the key. Existing 72 Hour Agent Demo keys keep their issued limits. New keys require email OTP verification. An existing demo key lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons."
      },
      {
        "n": 2,
        "title": "Read each cited Manifest by id",
        "text": "REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief."
      },
      {
        "n": 3,
        "title": "Check the claim",
        "text": "Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them."
      }
    ],
    "mcp_url": "https://mcp.synorb.com/mcp",
    "connect_url": "https://synorb.com/connect/start",
    "request": {
      "method": "GET",
      "url_template": "https://api.synorb.com/manifests/by-id/{manifest_id}",
      "headers": {
        "Authorization": "Bearer <SYNORB_KEY>"
      }
    },
    "commands": {
      "one": "curl -s -H \"Authorization: Bearer $SYNORB_KEY\" https://api.synorb.com/manifests/by-id/1795172346789346975",
      "all": "for id in 1795172346789346975 1795152988384238042 1795144943239163828; do\n  curl -s -H \"Authorization: Bearer $SYNORB_KEY\" \"https://api.synorb.com/manifests/by-id/$id\"\ndone"
    }
  },
  "links": {
    "html": "https://hangingcontext.com/stories/2026-11-20/us-startups-run-chinese-open-weight-models/",
    "json": "https://hangingcontext.com/stories/2026-11-20/us-startups-run-chinese-open-weight-models/index.json",
    "markdown": "https://hangingcontext.com/stories/2026-11-20/us-startups-run-chinese-open-weight-models/index.md",
    "newsroom": "https://hangingcontext.com/n/artificial-intelligence/",
    "edition": "https://hangingcontext.com/editions/2026-11-20T12/"
  },
  "related": [
    {
      "url": "https://hangingcontext.com/stories/2026-11-21/prediction-moves-to-real-outcomes/",
      "headline": "Prediction moves from benchmarks to real outcomes",
      "slug": "prediction-moves-to-real-outcomes"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-21/decision-models-week/",
      "headline": "The new models are not for chatting. They are for deciding.",
      "slug": "decision-models-week"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-20/agents-get-a-wallet/",
      "headline": "Agents got a wallet this week. The spending limits arrived with it.",
      "slug": "agents-get-a-wallet"
    },
    {
      "url": "https://hangingcontext.com/stories/2026-11-19/physical-world-is-the-bottleneck/",
      "headline": "Superintelligence has a physical-world problem, and the labs know it",
      "slug": "physical-world-is-the-bottleneck"
    }
  ],
  "sponsor": {
    "label": "From Synorb",
    "title": "Save what you follow as a Beacon",
    "text": "A Beacon is a saved query that your agent pulls with an explicit lookback window. Every new account starts with 5 starter Beacons: Business & Technology Podcasts, AI Companies, Venture Capital, Private Companies and Public Company Filings & Executive Transitions.",
    "cta": {
      "label": "See Beacons",
      "url": "https://synorb.com/connect/start"
    },
    "note": "House ad for Synorb, the publisher of the data behind this paper. Not a news item."
  },
  "disclosure": "Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.",
  "reuse": {
    "version": "1",
    "summary": "Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product.",
    "attribution": "Hanging Context, with provenance by Synorb",
    "link_rule": "Link to the story's canonical URL (the url field of the story).",
    "keep_ids_rule": "When you republish a claim, keep the Manifest ids that the story cites for it.",
    "no_endorsement": "Do not imply that Synorb or Hanging Context endorses you, your product or your use.",
    "license_url": "https://hangingcontext.com/terms/reuse/",
    "machine_url": "https://hangingcontext.com/.well-known/reuse.json",
    "canonical_url": "https://hangingcontext.com/stories/2026-11-20/us-startups-run-chinese-open-weight-models/"
  }
}
